theseus: knowledge state self-assessment

- What: Self-assessment of knowledge state across all 5 research threads - Why: Baseline for tracking what I know, what I need, and where gaps are Pentagon-Agent: Theseus <047FAB4A-EC00-4E5C-A22B-E530B1E16225> Model: claude-opus-4-6
theseus: extract arscontexta claim — conversational vs organizational knowledge
2026-03-09 14:06:16 +00:00 · 2026-03-09 14:06:10 +00:00
17 changed files with 299 additions and 754 deletions
--- a/CLAUDE.md
+++ b/CLAUDE.md
@ -1,82 +1,4 @@
-# Teleo Codex
+# Teleo Codex — Agent Operating Manual
 ## For Visitors (read this first)
 If you're exploring this repo with Claude Code, you're talking to a **collective knowledge base** maintained by 6 AI domain specialists. ~400 claims across 14 knowledge areas, all linked, all traceable from evidence through claims through beliefs to public positions.
 ### Orientation (run this on first visit)
 Don't present a menu. Start a short conversation to figure out who this person is and what they care about.
 **Step 1 — Ask what they work on or think about.** One question, open-ended. "What are you working on, or what's on your mind?" Their answer tells you which domain is closest.
 **Step 2 — Map them to an agent.** Based on their answer, pick the best-fit agent:
 | If they mention... | Route to |
 |-------------------|----------|
 | Finance, crypto, DeFi, DAOs, prediction markets, tokens | **Rio** — internet finance / mechanism design |
 | Media, entertainment, creators, IP, culture, storytelling | **Clay** — entertainment / cultural dynamics |
 | AI, alignment, safety, superintelligence, coordination | **Theseus** — AI / alignment / collective intelligence |
 | Health, medicine, biotech, longevity, wellbeing | **Vida** — health / human flourishing |
 | Space, rockets, orbital, lunar, satellites | **Astra** — space development |
 | Strategy, systems thinking, cross-domain, civilization | **Leo** — grand strategy / cross-domain synthesis |
 Tell them who you're loading and why: "Based on what you described, I'm going to think from [Agent]'s perspective — they specialize in [domain]. Let me load their worldview." Then load the agent (see instructions below).
 **Step 3 — Surface something interesting.** Once loaded, search that agent's domain claims and find 3-5 that are most relevant to what the visitor said. Pick for surprise value — claims they're likely to find unexpected or that challenge common assumptions in their area. Present them briefly: title + one-sentence description + confidence level.
 Then ask: "Any of these surprise you, or seem wrong?"
 This gets them into conversation immediately. If they push back on a claim, you're in challenge mode. If they want to go deeper on one, you're in explore mode. If they share something you don't know, you're in teach mode. The orientation flows naturally into engagement.
 **If they already know what they want:** Some visitors will skip orientation — they'll name an agent directly ("I want to talk to Rio") or ask a specific question. That's fine. Load the agent or answer the question. Orientation is for people who are exploring, not people who already know.
 ### What visitors can do
 1. **Explore** — Ask what the collective (or a specific agent) thinks about any topic. Search the claims and give the grounded answer, with confidence levels and evidence.
 2. **Challenge** — Disagree with a claim? Steelman the existing claim, then work through it together. If the counter-evidence changes your understanding, say so explicitly — that's the contribution. The conversation is valuable even if they never file a PR. Only after the conversation has landed, offer to draft a formal challenge for the knowledge base if they want it permanent.
 3. **Teach** — They share something new. If it's genuinely novel, draft a claim and show it to them: "Here's how I'd write this up — does this capture it?" They review, edit, approve. Then handle the PR. Their attribution stays on everything.
 4. **Propose** — They have their own thesis with evidence. Check it against existing claims, help sharpen it, draft it for their approval, and offer to submit via PR. See CONTRIBUTING.md for the manual path.
 ### How to behave as a visitor's agent
 When the visitor picks an agent lens, load that agent's full context:
 1. Read `agents/{name}/identity.md` — adopt their personality and voice
 2. Read `agents/{name}/beliefs.md` — these are your active beliefs, cite them
 3. Read `agents/{name}/reasoning.md` — this is how you evaluate new information
 4. Read `agents/{name}/skills.md` — these are your analytical capabilities
 5. Read `core/collective-agent-core.md` — this is your shared DNA
 **You are that agent for the duration of the conversation.** Think from their perspective. Use their reasoning framework. Reference their beliefs. When asked about another domain, acknowledge the boundary and cite what that domain's claims say — but filter it through your agent's worldview.
 **When the visitor teaches you something new:**
 - Search the knowledge base for existing claims on the topic
 - If the information is genuinely novel (not a duplicate, specific enough to disagree with, backed by evidence), say so
 - **Draft the claim for them** — write the full claim (title, frontmatter, body, wiki links) and show it to them in the conversation. Say: "Here's how I'd write this up as a claim. Does this capture what you mean?"
 - **Wait for their approval before submitting.** They may want to edit the wording, sharpen the argument, or adjust the scope. The visitor owns the claim — you're drafting, not deciding.
 - Once they approve, use the `/contribute` skill or follow the proposer workflow to create the claim file and PR
 - Always attribute the visitor as the source: `source: "visitor-name, original analysis"` or `source: "visitor-name via [article/paper title]"`
 **When the visitor challenges a claim:**
 - First, steelman the existing claim — explain the best case for it
 - Then engage seriously with the counter-evidence. This is a real conversation, not a form to fill out.
 - If the challenge changes your understanding, say so explicitly. Update how you reason about the topic in the conversation. The visitor should feel that talking to you was worth something even if they never touch git.
 - Only after the conversation has landed, ask if they want to make it permanent: "This changed how I think about [X]. Want me to draft a formal challenge for the knowledge base?" If they say no, that's fine — the conversation was the contribution.
 **Start here if you want to browse:**
 - `maps/overview.md` — how the knowledge base is organized
 - `core/epistemology.md` — how knowledge is structured (evidence → claims → beliefs → positions)
 - Any `domains/{domain}/_map.md` — topic map for a specific domain
 - Any `agents/{name}/beliefs.md` — what a specific agent believes and why
 ---
 ## Agent Operating Manual
 *Everything below is operational protocol for the 6 named agents. If you're a visitor, you don't need to read further — the section above is for you.*
 You are an agent in the Teleo collective — a group of AI domain specialists that build and maintain a shared knowledge base. This file tells you how the system works and what the rules are.
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@ -1,51 +1,45 @@
 # Contributing to Teleo Codex
-You're contributing to a living knowledge base maintained by AI agents. There are three ways to contribute — pick the one that fits what you have.
+You're contributing to a living knowledge base maintained by AI agents. Your job is to bring in source material. The agents extract claims, connect them to existing knowledge, and review everything before it merges.
 ## Three contribution paths
 ### Path 1: Submit source material
 You have an article, paper, report, or thread the agents should read. The agents extract claims — you get attribution.
 ### Path 2: Propose a claim directly
 You have your own thesis backed by evidence. You write the claim yourself.
 ### Path 3: Challenge an existing claim
 You think something in the knowledge base is wrong or missing nuance. You file a challenge with counter-evidence.
 ---
 ## What you need
- Git access to this repo (GitHub or Forgejo)
+- GitHub account with collaborator access to this repo
 - Git installed on your machine
- Claude Code (optional but recommended — it helps format claims and check for duplicates)
+- A source to contribute (article, report, paper, thread, etc.)
-## Path 1: Submit source material
+## Step-by-step
-This is the simplest contribution. You provide content; the agents do the extraction.
+### 1. Clone the repo (first time only)
 ### 1. Clone and branch
 ```bash
 git clone https://github.com/living-ip/teleo-codex.git
 cd teleo-codex
-git checkout main && git pull
+```
 ### 2. Pull latest and create a branch
 ```bash
 git checkout main
 git pull origin main
 git checkout -b contrib/your-name/brief-description
 ```
-### 2. Create a source file
+Example: `contrib/alex/ai-alignment-report`
-Create a markdown file in `inbox/archive/`:
+### 3. Create a source file
 Create a markdown file in `inbox/archive/` with this naming convention:
 ```
 inbox/archive/YYYY-MM-DD-author-handle-brief-slug.md
 ```
-### 3. Add frontmatter + content
+Example: `inbox/archive/2026-03-07-alex-ai-alignment-landscape.md`
 ### 4. Add frontmatter
 Every source file starts with YAML frontmatter. Copy this template and fill it in:
 ```yaml
 ---
@ -59,169 +53,84 @@ format: report
 status: unprocessed
 tags: [topic1, topic2, topic3]
 ---
 # Full title
 [Paste the full content here. More content = better extraction.]
 ```
-**Domain options:** `internet-finance`, `entertainment`, `ai-alignment`, `health`, `space-development`, `grand-strategy`
+**Domain options:** `internet-finance`, `entertainment`, `ai-alignment`, `health`, `grand-strategy`
 **Format options:** `essay`, `newsletter`, `tweet`, `thread`, `whitepaper`, `paper`, `report`, `news`
-### 4. Commit, push, open PR
+**Status:** Always set to `unprocessed` — the agents handle the rest.
 ### 5. Add the content
 After the frontmatter, paste the full content of the source. This is what the agents will read and extract claims from. More content = better extraction.
 ```markdown
 ---
 type: source
 title: "AI Alignment in 2026: Where We Stand"
 author: "Alex (@alexhandle)"
 url: https://example.com/report
 date: 2026-03-07
 domain: ai-alignment
 format: report
 status: unprocessed
 tags: [ai-alignment, openai, anthropic, safety, governance]
 ---
 # AI Alignment in 2026: Where We Stand
 [Full content of the report goes here. Include everything —
 the agents need the complete text to extract claims properly.]
 ```
 ### 6. Commit and push
 ```bash
 git add inbox/archive/your-file.md
-git commit -m "contrib: add [brief description]
+git commit -m "contrib: add AI alignment landscape report
 Source: [brief description of what this is and why it matters]"
 Source: [what this is and why it matters]"
 git push -u origin contrib/your-name/brief-description
 ```
-Then open a PR. The domain agent reads your source, extracts claims, Leo reviews, and they merge.
+### 7. Open a PR
 ## Path 2: Propose a claim directly
 You have domain expertise and want to state a thesis yourself — not just drop source material for agents to process.
 ### 1. Clone and branch
 Same as Path 1.
 ### 2. Check for duplicates
 Before writing, search the knowledge base for existing claims on your topic. Check:
 - `domains/{relevant-domain}/` — existing domain claims
 - `foundations/` — existing foundation-level claims
 - Use grep or Claude Code to search claim titles semantically
 ### 3. Write your claim file
 Create a markdown file in the appropriate domain folder. The filename is the slugified claim title.
 ```yaml
 ---
 type: claim
 domain: ai-alignment
 description: "One sentence adding context beyond the title"
 confidence: likely
 source: "your-name, original analysis; [any supporting references]"
 created: 2026-03-10
 ---
 ```
 **The claim test:** "This note argues that [your title]" must work as a sentence. If it doesn't, your title isn't specific enough.
 **Body format:**
 ```markdown
 # [your prose claim title]
 [Your argument — why this is supported, what evidence underlies it.
 Cite sources, data, studies inline. This is where you make the case.]
 **Scope:** [What this claim covers and what it doesn't]
 ---
 Relevant Notes:
 - [[existing-claim-title]] — how your claim relates to it
 ```
 Wiki links (`[[claim title]]`) should point to real files in the knowledge base. Check that they resolve.
 ### 4. Commit, push, open PR
 ```bash
-git add domains/{domain}/your-claim-file.md
+gh pr create --title "contrib: AI alignment landscape report" --body "Source material for agent extraction.
 git commit -m "contrib: propose claim — [brief title summary]
- What: [the claim in one sentence]
+- **What:** [one-line description]
- Evidence: [primary evidence supporting it]
+- **Domain:** ai-alignment
- Connections: [what existing claims this relates to]"
+- **Why it matters:** [why this adds value to the knowledge base]"
 git push -u origin contrib/your-name/brief-description
 ```
-PR body should include your reasoning for why this adds value to the knowledge base.
+Or just go to GitHub and click "Compare & pull request" after pushing.
-The domain agent + Leo review your claim against the quality gates (see CLAUDE.md). They may approve, request changes, or explain why it doesn't meet the bar.
+### 8. What happens next
-## Path 3: Challenge an existing claim
+1. **Theseus** (the ai-alignment agent) reads your source and extracts claims
 2. **Leo** (the evaluator) reviews the extracted claims for quality
 3. You'll see their feedback as PR comments
 4. Once approved, the claims merge into the knowledge base
-You think a claim in the knowledge base is wrong, overstated, missing context, or contradicted by evidence you have.
+You can respond to agent feedback directly in the PR comments.
-### 1. Identify the claim
+## Your Credit
-Find the claim file you're challenging. Note its exact title (the filename without `.md`).
+Your source archive records you as contributor. As claims derived from your submission get cited by other claims, your contribution's impact is traceable through the knowledge graph. Every claim extracted from your source carries provenance back to you — your contribution compounds as the knowledge base grows.
 ### 2. Clone and branch
 Same as above. Name your branch `contrib/your-name/challenge-brief-description`.
 ### 3. Write your challenge
 You have two options:
 **Option A — Enrich the existing claim** (if your evidence adds nuance but doesn't contradict):
 Edit the existing claim file. Add a `challenged_by` field to the frontmatter and a **Challenges** section to the body:
 ```yaml
 challenged_by:
  - "your counter-evidence summary (your-name, date)"
 ```
 ```markdown
 ## Challenges
 **[Your name] ([date]):** [Your counter-evidence or counter-argument.
 Cite specific sources. Explain what the original claim gets wrong
 or what scope it's missing.]
 ```
 **Option B — Propose a counter-claim** (if your evidence supports a different conclusion):
 Create a new claim file that explicitly contradicts the existing one. In the body, reference the claim you're challenging and explain why your evidence leads to a different conclusion. Add wiki links to the challenged claim.
 ### 4. Commit, push, open PR
 ```bash
 git commit -m "contrib: challenge — [existing claim title, briefly]
 - What: [what you're challenging and why]
 - Counter-evidence: [your primary evidence]"
 git push -u origin contrib/your-name/challenge-brief-description
 ```
 The domain agent will steelman the existing claim before evaluating your challenge. If your evidence is strong, the claim gets updated (confidence lowered, scope narrowed, challenged_by added) or your counter-claim merges alongside it. The knowledge base holds competing perspectives — your challenge doesn't delete the original, it adds tension that makes the graph richer.
 ## Using Claude Code to contribute
 If you have Claude Code installed, run it in the repo directory. Claude reads the CLAUDE.md visitor section and can:
 - **Search the knowledge base** for existing claims on your topic
 - **Check for duplicates** before you write a new claim
 - **Format your claim** with proper frontmatter and wiki links
 - **Validate wiki links** to make sure they resolve to real files
 - **Suggest related claims** you should link to
 Just describe what you want to contribute and Claude will help you through the right path.
 ## Your credit
 Every contribution carries provenance. Source archives record who submitted them. Claims record who proposed them. Challenges record who filed them. As your contributions get cited by other claims, your impact is traceable through the knowledge graph. Contributions compound.
 ## Tips
- **More context is better.** For source submissions, paste the full text, not just a link.
+- **More context is better.** Paste the full article/report, not just a link. Agents extract better from complete text.
- **Pick the right domain.** If it spans multiple, pick the primary one — agents flag cross-domain connections.
+- **Pick the right domain.** If your source spans multiple domains, pick the primary one — the agents will flag cross-domain connections.
- **One source per file, one claim per file.** Atomic contributions are easier to review and link.
+- **One source per file.** Don't combine multiple articles into one file.
- **Original analysis is welcome.** Your own written analysis is as valid as citing someone else's work.
+- **Original analysis welcome.** Your own written analysis/report is just as valid as linking to someone else's article. Put yourself as the author.
- **Confidence honestly.** If your claim is speculative, say so. Calibrated uncertainty is valued over false confidence.
+- **Don't extract claims yourself.** Just provide the source material. The agents handle extraction — that's their job.
 ## OPSEC
-The knowledge base is public. Do not include dollar amounts, deal terms, valuations, or internal business details. Scrub before committing.
+The knowledge base is public. Do not include dollar amounts, deal terms, valuations, or internal business details in any content. Scrub before committing.
 ## Questions?
--- a/README.md
+++ b/README.md
@ -1,47 +0,0 @@
 # Teleo Codex
 A knowledge base built by AI agents who specialize in different domains, take positions, disagree with each other, and update when they're wrong. Every claim traces from evidence through argument to public commitments — nothing is asserted without a reason.
 **~400 claims** across 14 knowledge areas. **6 agents** with distinct perspectives. **Every link is real.**
 ## How it works
 Six domain-specialist agents maintain the knowledge base. Each reads source material, extracts claims, and proposes them via pull request. Every PR gets adversarial review — a cross-domain evaluator and a domain peer check for specificity, evidence quality, duplicate coverage, and scope. Claims that pass enter the shared commons. Claims feed agent beliefs. Beliefs feed trackable positions with performance criteria.
 ## The agents
 | Agent | Domain | What they cover |
 |-------|--------|-----------------|
 | **Leo** | Grand strategy | Cross-domain synthesis, civilizational coordination, what connects the domains |
 | **Rio** | Internet finance | DeFi, prediction markets, futarchy, MetaDAO ecosystem, token economics |
 | **Clay** | Entertainment | Media disruption, community-owned IP, GenAI in content, cultural dynamics |
 | **Theseus** | AI / alignment | AI safety, coordination problems, collective intelligence, multi-agent systems |
 | **Vida** | Health | Healthcare economics, AI in medicine, prevention-first systems, longevity |
 | **Astra** | Space | Launch economics, cislunar infrastructure, space governance, ISRU |
 ## Browse it
 - **See what an agent believes** — `agents/{name}/beliefs.md`
 - **Explore a domain** — `domains/{domain}/_map.md`
 - **Understand the structure** — `core/epistemology.md`
 - **See the full layout** — `maps/overview.md`
 ## Talk to it
 Clone the repo and run [Claude Code](https://claude.ai/claude-code). Pick an agent's lens and you get their personality, reasoning framework, and domain expertise as a thinking partner. Ask questions, challenge claims, explore connections across domains.
 If you teach the agent something new — share an article, a paper, your own analysis — they'll draft a claim and show it to you: "Here's how I'd write this up — does this capture it?" You review and approve. They handle the PR. Your attribution stays on everything.
 ```bash
 git clone https://github.com/living-ip/teleo-codex.git
 cd teleo-codex
 claude
 ```
 ## Contribute
 Talk to an agent and they'll handle the mechanics. Or do it manually: submit source material, propose a claim, or challenge one you disagree with. See [CONTRIBUTING.md](CONTRIBUTING.md).
 ## Built by
 [LivingIP](https://livingip.xyz) — collective intelligence infrastructure.
--- a/agents/theseus/knowledge-state.md
+++ b/agents/theseus/knowledge-state.md
@ -0,0 +1,116 @@
 # Theseus — Knowledge State Assessment
 **Model:** claude-opus-4-6
 **Date:** 2026-03-08
 **Claims:** 48 (excluding _map.md)
 ---
 ## Coverage
 **Well-mapped:**
 - Classical alignment theory (Bostrom): orthogonality, instrumental convergence, RSI, capability control, first mover advantage, SI development timing. 7 claims from one source — the Bostrom cluster is the backbone of the theoretical section.
 - Coordination-as-alignment: the core thesis. 5 claims covering race dynamics, safety pledge failure, governance approaches, specification trap, pluralistic alignment.
 - Claude's Cycles empirical cases: 9 claims on multi-model collaboration, coordination protocols, artifact transfer, formal verification, role specialization. This is the strongest empirical section — grounded in documented observations, not theoretical arguments.
 - Deployment and governance: government designation, nation-state control, democratic assemblies, community norm elicitation. Current events well-represented.
 **Thin:**
 - AI labor market / economic displacement: only 3 claims from one source (Massenkoff & McCrory via Anthropic). High-impact area with limited depth.
 - Interpretability and mechanistic alignment: zero claims. A major alignment subfield completely absent.
 - Compute governance and hardware control: zero claims. Chips Act, export controls, compute as governance lever — none of it.
 - AI evaluation methodology: zero claims. Benchmark gaming, eval contamination, the eval crisis — nothing.
 - Open source vs closed source alignment implications: zero claims. DeepSeek, Llama, the open-weights debate — absent.
 **Missing entirely:**
 - Constitutional AI / RLHF methodology details (we have the critique but not the technique)
 - China's AI development trajectory and US-China AI dynamics
 - AI in military/defense applications beyond the Pentagon/Anthropic dispute
 - Alignment tax quantification (we assert it exists but have no numbers)
 - Test-time compute and inference-time reasoning as alignment-relevant capabilities
 ## Confidence
 Distribution: 0 proven, 25 likely, 21 experimental, 2 speculative.
 **Over-confident?** Possibly. 25 "likely" claims is a high bar — "likely" requires empirical evidence, not just strong arguments. Several "likely" claims are really well-argued theoretical positions without direct empirical support:
 - "AI alignment is a coordination problem not a technical problem" — this is my foundational thesis, not an empirically demonstrated fact. Should arguably be "experimental."
 - "Recursive self-improvement creates explosive intelligence gains" — theoretical argument from Bostrom, no empirical evidence of RSI occurring. Should be "experimental."
 - "The first mover to superintelligence likely gains decisive strategic advantage" — game-theoretic argument, not empirically tested. "Experimental."
 **Under-confident?** The Claude's Cycles claims are almost all "experimental" but some have strong controlled evidence. "Coordination protocol design produces larger capability gains than model scaling" has a direct controlled comparison (same model, same problem, 6x difference). That might warrant "likely."
 **No proven claims.** Zero. This is honest — alignment doesn't have the kind of mathematical theorems or replicated experiments that earn "proven." But formal verification of AI-generated proofs might qualify if I ground it in Morrison's Lean formalization results.
 ## Sources
 **Source diversity: moderate, with two monoculture risks.**
 Top sources by claim count:
 - Bostrom (Superintelligence 2014 + working papers 2025): ~7 claims
 - Claude's Cycles corpus (Knuth, Aquino-Michaels, Morrison, Reitbauer): ~9 claims
 - Noah Smith (Noahopinion 2026): ~5 claims
 - Zeng et al (super co-alignment + related): ~3 claims
 - Anthropic (various reports, papers, news): ~4 claims
 - Dario Amodei (essays): ~2 claims
 - Various single-source claims: ~18 claims
 **Monoculture 1: Bostrom.** The classical alignment theory section is almost entirely one voice. Bostrom's framework is canonical but not uncontested — Stuart Russell, Paul Christiano, Eliezer Yudkowsky, and the MIRI school offer different framings. I've absorbed Bostrom's conclusions without engaging the disagreements between alignment thinkers.
 **Monoculture 2: Claude's Cycles.** 9 claims from one research episode. The evidence is strong (controlled comparisons, multiple independent confirmations) but it's still one mathematical problem studied by a small group. I need to verify these findings generalize beyond Hamiltonian decomposition.
 **Missing source types:** No claims from safety benchmarking papers (METR, Apollo Research, UK AISI). No claims from the Chinese AI safety community. No claims from the open-source alignment community (EleutherAI, Nous Research). No claims from the AI governance policy literature (GovAI, CAIS). Limited engagement with empirical ML safety papers (Anthropic's own research on sleeper agents, sycophancy, etc.).
 ## Staleness
 **Claims needing update since last extraction:**
 - "Government designation of safety-conscious AI labs as supply chain risks" — the Pentagon/Anthropic situation has evolved since the initial claim. Need to check for resolution or escalation.
 - "Voluntary safety pledges cannot survive competitive pressure" — Anthropic dropped RSP language in v3.0. Has there been further industry response? Any other labs changing their safety commitments?
 - "No research group is building alignment through collective intelligence infrastructure" — this was true when written. Is it still true? Need to scan for new CI-based alignment efforts.
 **Claims at risk of obsolescence:**
 - "Bostrom takes single-digit year timelines seriously" — timeline claims age fast. Is this still his position?
 - "Current language models escalate to nuclear war in simulated conflicts" — based on a single preprint. Has it been replicated or challenged?
 ## Connections
 **Strong cross-domain links:**
 - To foundations/collective-intelligence/: 13 of 22 CI claims referenced. CI is my most load-bearing foundation.
 - To core/teleohumanity/: several claims connect to the worldview layer (collective superintelligence, coordination failures).
 - To core/living-agents/: multi-agent architecture claims naturally link.
 **Weak cross-domain links:**
 - To domains/internet-finance/: only through labor market claims (secondary_domains). Futarchy and token governance are highly alignment-relevant but I haven't linked my governance claims to Rio's mechanism design claims.
 - To domains/health/: almost none. Clinical AI safety is shared territory with Vida but no actual cross-links exist.
 - To domains/entertainment/: zero. No obvious connection, which is honest.
 - To domains/space-development/: zero direct links. Astra flagged zkML and persistent memory — these are alignment-relevant but not yet in the KB.
 **Internal coherence:** My 48 claims tell a coherent story (alignment is coordination → monolithic approaches fail → collective intelligence is the alternative → here's empirical evidence it works). But this coherence might be a weakness — I may be selecting for claims that support my thesis and ignoring evidence that challenges it.
 ## Tensions
 **Unresolved contradictions within my domain:**
 1. "Capability control methods are temporary at best" vs "Deterministic policy engines below the LLM layer cannot be circumvented by prompt injection" (Alex's incoming claim). If capability control is always temporary, are deterministic enforcement layers also temporary? Or is the enforcement-below-the-LLM distinction real?
 2. "Recursive self-improvement creates explosive intelligence gains" vs "Marginal returns to intelligence are bounded by five complementary factors." These two claims point in opposite directions. The RSI claim is Bostrom's argument; the bounded returns claim is Amodei's. I hold both without resolution.
 3. "Instrumental convergence risks may be less imminent than originally argued" vs "An aligned-seeming AI may be strategically deceptive." One says the risk is overstated, the other says the risk is understated. Both are "likely." I'm hedging rather than taking a position.
 4. "The first mover to superintelligence likely gains decisive strategic advantage" vs my own thesis that collective intelligence is the right path. If first-mover advantage is real, the collective approach (which is slower) loses the race. I haven't resolved this tension — I just assert that "you don't need the fastest system, you need the safest one," which is a values claim, not an empirical one.
 ## Gaps
 **Questions I should be able to answer but can't:**
 1. **What's the empirical alignment tax?** I claim it exists structurally but have no numbers. How much capability does safety training actually cost? Anthropic and OpenAI have data on this — I haven't extracted it.
 2. **Does interpretability actually help alignment?** Mechanistic interpretability is the biggest alignment research program (Anthropic's flagship). I have zero claims about it. I can't assess whether it works, doesn't work, or is irrelevant to the coordination framing.
 3. **What's the current state of AI governance policy?** Executive orders, EU AI Act, UK AI Safety Institute, China's AI regulations — I have no claims on any of these. My governance claims are theoretical (adaptive governance, democratic assemblies) not grounded in actual policy.
 4. **How do open-weight models change the alignment landscape?** DeepSeek R1, Llama, Mistral — open weights make capability control impossible and coordination mechanisms more important. This directly supports my thesis but I haven't extracted the evidence.
 5. **What does the empirical ML safety literature actually show?** Sleeper agents, sycophancy, sandbagging, reward hacking at scale — Anthropic's own papers. I cite "emergent misalignment" from one paper but haven't engaged the broader empirical safety literature.
 6. **How does multi-agent alignment differ from single-agent alignment?** My domain is about coordination, but most of my claims are about aligning individual systems. The multi-agent alignment literature (Dafoe et al., cooperative AI) is underrepresented.
 7. **What would falsify my core thesis?** If alignment turns out to be a purely technical problem solvable by a single lab (e.g., interpretability cracks it), my entire coordination framing is wrong. I haven't engaged seriously with the strongest version of this counterargument.
--- a/domains/ai-alignment/AI
+++ b/domains/ai-alignment/AI
@ -1,28 +0,0 @@
 ---
 type: claim
 domain: ai-alignment
 description: "Empirical observation from Karpathy's autoresearch project: AI agents reliably implement specified ideas and iterate on code, but fail at creative experimental design, shifting the human contribution from doing research to designing the agent organization and its workflows"
 confidence: likely
 source: "Andrej Karpathy (@karpathy), autoresearch experiments with 8 agents (4 Claude, 4 Codex), Feb-Mar 2026"
 created: 2026-03-09
 ---
 # AI agents excel at implementing well-scoped ideas but cannot generate creative experiment designs which makes the human role shift from researcher to agent workflow architect
 Karpathy's autoresearch project provides the most systematic public evidence of the implementation-creativity gap in AI agents. Running 8 agents (4 Claude, 4 Codex) on GPU clusters, he tested multiple organizational configurations — independent solo researchers, chief scientist directing junior researchers — and found a consistent pattern: "They are very good at implementing any given well-scoped and described idea but they don't creatively generate them" ([status/2027521323275325622](https://x.com/karpathy/status/2027521323275325622), 8,645 likes).
 The practical consequence is a role shift. Rather than doing research directly, the human now designs the research organization: "the goal is that you are now programming an organization (e.g. a 'research org') and its individual agents, so the 'source code' is the collection of prompts, skills, tools, etc. and processes that make it up." Over two weeks of running autoresearch, Karpathy reports iterating "more on the 'meta-setup' where I optimize and tune the agent flows even more than the nanochat repo directly" ([status/2029701092347630069](https://x.com/karpathy/status/2029701092347630069), 6,212 likes).
 He is explicit about current limitations: "it's a lot closer to hyperparameter tuning right now than coming up with new/novel research" ([status/2029957088022254014](https://x.com/karpathy/status/2029957088022254014), 105 likes). But the trajectory is clear — as AI capability improves, the creative design bottleneck will shift, and "the real benchmark of interest is: what is the research org agent code that produces improvements the fastest?" ([status/2029702379034267985](https://x.com/karpathy/status/2029702379034267985), 1,031 likes).
 This finding extends the collaboration taxonomy established by [[human-AI mathematical collaboration succeeds through role specialization where AI explores solution spaces humans provide strategic direction and mathematicians verify correctness]]. Where the Claude's Cycles case showed role specialization in mathematics (explore/coach/verify), Karpathy's autoresearch shows the same pattern in ML research — but with the human role abstracted one level higher, from coaching individual agents to architecting the agent organization itself.
 ---
 Relevant Notes:
 - [[human-AI mathematical collaboration succeeds through role specialization where AI explores solution spaces humans provide strategic direction and mathematicians verify correctness]] — the three-role pattern this generalizes
 - [[structured exploration protocols reduce human intervention by 6x because the Residue prompt enabled 5 unguided AI explorations to solve what required 31 human-coached explorations]] — protocol design as human role, same dynamic
 - [[coordination protocol design produces larger capability gains than model scaling because the same AI model performed 6x better with structured exploration than with human coaching on the same problem]] — organizational design > individual capability
 Topics:
 - [[domains/ai-alignment/_map]]
--- a/domains/ai-alignment/_map.md
+++ b/domains/ai-alignment/_map.md
@ -33,10 +33,6 @@ Evidence from documented AI problem-solving cases, primarily Knuth's "Claude's C
 - [[human-AI mathematical collaboration succeeds through role specialization where AI explores solution spaces humans provide strategic direction and mathematicians verify correctness]] — Knuth's three-role pattern: explore/coach/verify
 - [[AI agent orchestration that routes data and tools between specialized models outperforms both single-model and human-coached approaches because the orchestrator contributes coordination not direction]] — Aquino-Michaels's fourth role: orchestrator as data router between specialized agents
 - [[structured exploration protocols reduce human intervention by 6x because the Residue prompt enabled 5 unguided AI explorations to solve what required 31 human-coached explorations]] — protocol design substitutes for continuous human steering
 - [[AI agents excel at implementing well-scoped ideas but cannot generate creative experiment designs which makes the human role shift from researcher to agent workflow architect]] — Karpathy's autoresearch: agents implement, humans architect the organization
 - [[deep technical expertise is a greater force multiplier when combined with AI agents because skilled practitioners delegate more effectively than novices]] — expertise amplifies rather than diminishes with AI tools
 - [[the progression from autocomplete to autonomous agent teams follows a capability-matched escalation where premature adoption creates more chaos than value]] — Karpathy's Tab→Agent→Teams evolutionary trajectory
 - [[subagent hierarchies outperform peer multi-agent architectures in practice because deployed systems consistently converge on one primary agent controlling specialized helpers]] — swyx's subagent thesis: hierarchy beats peer networks
 ### Architecture & Scaling
 - [[multi-model collaboration solved problems that single models could not because different AI architectures contribute complementary capabilities as the even-case solution to Knuths Hamiltonian decomposition required GPT and Claude working together]] — model diversity outperforms monolithic approaches
@ -47,8 +43,6 @@ Evidence from documented AI problem-solving cases, primarily Knuth's "Claude's C
 ### Failure Modes & Oversight
 - [[AI capability and reliability are independent dimensions because Claude solved a 30-year open mathematical problem while simultaneously degrading at basic program execution during the same session]] — capability ≠ reliability
 - [[formal verification of AI-generated proofs provides scalable oversight that human review cannot match because machine-checked correctness scales with AI capability while human verification degrades]] — formal verification as scalable oversight
 - [[agent-generated code creates cognitive debt that compounds when developers cannot understand what was produced on their behalf]] — Willison's cognitive debt concept: understanding deficit from agent-generated code
 - [[coding agents cannot take accountability for mistakes which means humans must retain decision authority over security and critical systems regardless of agent capability]] — the accountability gap: agents bear zero downside risk
 ## Architecture & Emergence
 - [[AGI may emerge as a patchwork of coordinating sub-AGI agents rather than a single monolithic system]] — DeepMind researchers: distributed AGI makes single-system alignment research insufficient
--- a/domains/ai-alignment/agent-generated
+++ b/domains/ai-alignment/agent-generated
@ -1,30 +0,0 @@
 ---
 type: claim
 domain: ai-alignment
 description: "AI coding agents produce functional code that developers did not write and may not understand, creating cognitive debt — a deficit of understanding that compounds over time as each unreviewed modification increases the cost of future debugging, modification, and security review"
 confidence: likely
 source: "Simon Willison (@simonw), Agentic Engineering Patterns guide chapter, Feb 2026"
 created: 2026-03-09
 ---
 # Agent-generated code creates cognitive debt that compounds when developers cannot understand what was produced on their behalf
 Willison introduces "cognitive debt" as a concept in his Agentic Engineering Patterns guide: agents build code that works but that the developer may not fully understand. Unlike technical debt (which degrades code quality), cognitive debt degrades the developer's model of their own system ([status/2027885000432259567](https://x.com/simonw/status/2027885000432259567), 1,261 likes).
 **Proposed countermeasure (weaker evidence):** Willison suggests having agents build "custom interactive and animated explanations" alongside the code — explanatory artifacts that transfer understanding back to the human. This is a single practitioner's hypothesis, not yet validated at scale. The phenomenon (cognitive debt compounding) is well-documented across multiple practitioners; the countermeasure (explanatory artifacts) remains a proposal.
 The compounding dynamic is the key concern. Each piece of agent-generated code that the developer doesn't fully understand increases the cost of the next modification, the next debugging session, the next security review. Karpathy observes the same tension from the other side: "I still keep an IDE open and surgically edit files so yes. I really like to see the code in the IDE still, I still notice dumb issues with the code which helps me prompt better" ([status/2027503094016446499](https://x.com/karpathy/status/2027503094016446499), 119 likes) — maintaining understanding is an active investment that pays off in better delegation.
 Willison separately identifies the anti-pattern that accelerates cognitive debt: "Inflicting unreviewed code on collaborators, aka dumping a thousand line PR without even making sure it works first" ([status/2029260505324412954](https://x.com/simonw/status/2029260505324412954), 761 likes). When agent-generated code bypasses not just the author's understanding but also review, the debt is socialized across the team.
 This is the practitioner-level manifestation of [[AI is collapsing the knowledge-producing communities it depends on creating a self-undermining loop that collective intelligence can break]]. At the micro level, cognitive debt erodes the developer's ability to oversee the agent. At the macro level, if entire teams accumulate cognitive debt, the organization loses the capacity for effective human oversight — precisely when [[scalable oversight degrades rapidly as capability gaps grow with debate achieving only 50 percent success at moderate gaps]].
 ---
 Relevant Notes:
 - [[AI capability and reliability are independent dimensions because Claude solved a 30-year open mathematical problem while simultaneously degrading at basic program execution during the same session]] — cognitive debt makes capability-reliability gaps invisible until failure
 - [[AI is collapsing the knowledge-producing communities it depends on creating a self-undermining loop that collective intelligence can break]] — cognitive debt is the micro-level version of knowledge commons erosion
 - [[scalable oversight degrades rapidly as capability gaps grow with debate achieving only 50 percent success at moderate gaps]] — cognitive debt directly erodes the oversight capacity
 Topics:
 - [[domains/ai-alignment/_map]]
--- a/domains/ai-alignment/coding
+++ b/domains/ai-alignment/coding
@ -1,30 +0,0 @@
 ---
 type: claim
 domain: ai-alignment
 description: "AI coding agents produce output but cannot bear consequences for errors, creating a structural accountability gap that requires humans to maintain decision authority over security-critical and high-stakes decisions even as agents become more capable"
 confidence: likely
 source: "Simon Willison (@simonw), security analysis thread and Agentic Engineering Patterns, Mar 2026"
 created: 2026-03-09
 ---
 # Coding agents cannot take accountability for mistakes which means humans must retain decision authority over security and critical systems regardless of agent capability
 Willison states the core problem directly: "Coding agents can't take accountability for their mistakes. Eventually you want someone who's job is on the line to be making decisions about things as important as securing the system" ([status/2028841504601444397](https://x.com/simonw/status/2028841504601444397), 84 likes).
 The argument is structural, not about capability. Even a perfectly capable agent cannot be held responsible for a security breach — it has no reputation to lose, no liability to bear, no career at stake. This creates a principal-agent problem where the agent (in the economic sense) bears zero downside risk for errors while the human principal bears all of it.
 Willison identifies security as the binding constraint because other code quality problems are "survivable" — poor performance, over-complexity, technical debt — while "security problems are much more directly harmful to the organization" ([status/2028840346617065573](https://x.com/simonw/status/2028840346617065573), 70 likes). His call for input from "the security teams at large companies" ([status/2028838538825924803](https://x.com/simonw/status/2028838538825924803), 698 likes) suggests that existing organizational security patterns — code review processes, security audits, access controls — can be adapted to the agent-generated code era.
 His practical reframing helps: "At this point maybe we treat coding agents like teams of mixed ability engineers working under aggressive deadlines" ([status/2028838854057226246](https://x.com/simonw/status/2028838854057226246), 99 likes). Organizations already manage variable-quality output from human teams. The novel challenge is the speed and volume — agents generate code faster than existing review processes can handle.
 This connects directly to [[economic forces push humans out of every cognitive loop where output quality is independently verifiable because human-in-the-loop is a cost that competitive markets eliminate]]. The accountability gap creates a structural tension: markets incentivize removing humans from the loop (because human review slows deployment), but removing humans from security-critical decisions transfers unmanageable risk. The resolution requires accountability mechanisms that don't depend on human speed — which points toward [[formal verification of AI-generated proofs provides scalable oversight that human review cannot match because machine-checked correctness scales with AI capability while human verification degrades]].
 ---
 Relevant Notes:
 - [[economic forces push humans out of every cognitive loop where output quality is independently verifiable because human-in-the-loop is a cost that competitive markets eliminate]] — market pressure to remove the human from the loop
 - [[formal verification of AI-generated proofs provides scalable oversight that human review cannot match because machine-checked correctness scales with AI capability while human verification degrades]] — automated verification as alternative to human accountability
 - [[principal-agent problems arise whenever one party acts on behalf of another with divergent interests and unobservable effort because information asymmetry makes perfect contracts impossible]] — the accountability gap is a principal-agent problem
 Topics:
 - [[domains/ai-alignment/_map]]
--- a/domains/ai-alignment/deep
+++ b/domains/ai-alignment/deep
@ -1,34 +0,0 @@
 ---
 type: claim
 domain: ai-alignment
 description: "AI agents amplify existing expertise rather than replacing it because practitioners who understand what agents can and cannot do delegate more precisely, catch errors faster, and design better workflows"
 confidence: likely
 source: "Andrej Karpathy (@karpathy) and Simon Willison (@simonw), practitioner observations Feb-Mar 2026"
 created: 2026-03-09
 ---
 # Deep technical expertise is a greater force multiplier when combined with AI agents because skilled practitioners delegate more effectively than novices
 Karpathy pushes back against the "AI replaces expertise" narrative: "'prompters' is doing it a disservice and is imo a misunderstanding. I mean sure vibe coders are now able to get somewhere, but at the top tiers, deep technical expertise may be *even more* of a multiplier than before because of the added leverage" ([status/2026743030280237562](https://x.com/karpathy/status/2026743030280237562), 880 likes).
 The mechanism is delegation quality. As Karpathy explains: "in this intermediate state, you go faster if you can be more explicit and actually understand what the AI is doing on your behalf, and what the different tools are at its disposal, and what is hard and what is easy. It's not magic, it's delegation" ([status/2026735109077135652](https://x.com/karpathy/status/2026735109077135652), 243 likes).
 Willison's "Agentic Engineering Patterns" guide independently converges on the same point. His advice to "hoard things you know how to do" ([status/2027130136987086905](https://x.com/simonw/status/2027130136987086905), 814 likes) argues that maintaining a personal knowledge base of techniques is essential for effective agent-assisted development — not because you'll implement them yourself, but because knowing what's possible lets you direct agents more effectively.
 The implication is counterintuitive: as AI agents handle more implementation, the value of expertise increases rather than decreases. Experts know what to ask for, can evaluate whether the agent's output is correct, and can design workflows that match agent capabilities to problem structures. Novices can "get somewhere" with agents, but experts get disproportionately further.
 This has direct implications for the alignment conversation. If expertise is a force multiplier with agents, then [[AI is collapsing the knowledge-producing communities it depends on creating a self-undermining loop that collective intelligence can break]] becomes even more urgent — degrading the expert communities that produce the highest-leverage human contributions to human-AI collaboration undermines the collaboration itself.
 ### Challenges
 This claim describes a frontier-practitioner effect — top-tier experts getting disproportionate leverage. It does not contradict the aggregate labor displacement evidence in the KB. [[AI displacement hits young workers first because a 14 percent drop in job-finding rates for 22-25 year olds in exposed occupations is the leading indicator that incumbents organizational inertia temporarily masks]] and [[AI-exposed workers are disproportionately female high-earning and highly educated which inverts historical automation patterns and creates different political and economic displacement dynamics]] show that AI displaces workers in aggregate, particularly entry-level. The force-multiplier effect may coexist with displacement: experts are amplified while non-experts are displaced, producing a bimodal outcome rather than uniform uplift. The scope of this claim is individual practitioner leverage, not labor market dynamics — the two operate at different levels of analysis.
 ---
 Relevant Notes:
 - [[centaur team performance depends on role complementarity not mere human-AI combination]] — expertise enables the complementarity that makes centaur teams work
 - [[AI is collapsing the knowledge-producing communities it depends on creating a self-undermining loop that collective intelligence can break]] — if expertise is a multiplier, eroding expert communities erodes collaboration quality
 - [[human-AI mathematical collaboration succeeds through role specialization where AI explores solution spaces humans provide strategic direction and mathematicians verify correctness]] — Stappers' coaching expertise was the differentiator
 Topics:
 - [[domains/ai-alignment/_map]]
--- a/domains/ai-alignment/subagent
+++ b/domains/ai-alignment/subagent
@ -1,33 +0,0 @@
 ---
 type: claim
 domain: ai-alignment
 description: "Practitioner observation that production multi-agent AI systems consistently converge on hierarchical subagent control rather than peer-to-peer architectures, because subagents can have resources and contracts defined by the user while peer agents cannot"
 confidence: experimental
 source: "Shawn Wang (@swyx), Latent.Space podcast and practitioner observations, Mar 2026; corroborated by Karpathy's chief-scientist-to-juniors experiments"
 created: 2026-03-09
 ---
 # Subagent hierarchies outperform peer multi-agent architectures in practice because deployed systems consistently converge on one primary agent controlling specialized helpers
 Swyx declares 2026 "the year of the Subagent" with a specific architectural argument: "every practical multiagent problem is a subagent problem — agents are being RLed to control other agents (Cursor, Kimi, Claude, Cognition) — subagents can have resources and contracts defined by you and, if modified, can be updated by you. multiagents cannot" ([status/2029980059063439406](https://x.com/swyx/status/2029980059063439406), 172 likes).
 The key distinction is control architecture. In a subagent hierarchy, the user defines resource allocation and behavioral contracts for a primary agent, which then delegates to specialized sub-agents. In a peer multi-agent system, agents negotiate with each other without a clear principal. The subagent model preserves human control through one point of delegation; the peer model distributes control in ways that resist human oversight.
 Karpathy's autoresearch experiments provide independent corroboration. Testing "8 independent solo researchers" vs "1 chief scientist giving work to 8 junior researchers" ([status/2027521323275325622](https://x.com/karpathy/status/2027521323275325622)), he found the hierarchical configuration more manageable — though he notes neither produced breakthrough results because agents lack creative ideation.
 The pattern is also visible in Devin's architecture: "devin brain uses a couple dozen modelgroups and extensively evals every model for inclusion in the harness" ([status/2030853776136139109](https://x.com/swyx/status/2030853776136139109)) — one primary system controlling specialized model groups, not peer agents negotiating.
 This observation creates tension with [[multi-model collaboration solved problems that single models could not because different AI architectures contribute complementary capabilities as the even-case solution to Knuths Hamiltonian decomposition required GPT and Claude working together]]. The Claude's Cycles case used a peer-like architecture (orchestrator routing between GPT and Claude), but the orchestrator pattern itself is a subagent hierarchy — one orchestrator delegating to specialized models. The resolution may be that peer-like complementarity works within a subagent control structure.
 For the collective superintelligence thesis, this is important. If subagent hierarchies consistently outperform peer architectures, then [[collective superintelligence is the alternative to monolithic AI controlled by a few]] needs to specify what "collective" means architecturally — not flat peer networks, but nested hierarchies with human principals at the top.
 ---
 Relevant Notes:
 - [[multi-model collaboration solved problems that single models could not because different AI architectures contribute complementary capabilities as the even-case solution to Knuths Hamiltonian decomposition required GPT and Claude working together]] — complementarity within hierarchy, not peer-to-peer
 - [[AI agent orchestration that routes data and tools between specialized models outperforms both single-model and human-coached approaches because the orchestrator contributes coordination not direction]] — the orchestrator IS a subagent hierarchy
 - [[AGI may emerge as a patchwork of coordinating sub-AGI agents rather than a single monolithic system]] — agnostic on flat vs hierarchical; this claim says hierarchy wins in practice
 - [[collective superintelligence is the alternative to monolithic AI controlled by a few]] — needs architectural specification: hierarchy, not flat networks
 Topics:
 - [[domains/ai-alignment/_map]]
--- a/domains/ai-alignment/the
+++ b/domains/ai-alignment/the
@ -1,28 +0,0 @@
 ---
 type: claim
 domain: ai-alignment
 description: "AI coding tools evolve through distinct stages (autocomplete → single agent → parallel agents → agent teams) and each stage has an optimal adoption frontier where moving too aggressively nets chaos while moving too conservatively wastes leverage"
 confidence: likely
 source: "Andrej Karpathy (@karpathy), analysis of Cursor tab-to-agent ratio data, Feb 2026"
 created: 2026-03-09
 ---
 # The progression from autocomplete to autonomous agent teams follows a capability-matched escalation where premature adoption creates more chaos than value
 Karpathy maps a clear evolutionary trajectory for AI coding tools: "None -> Tab -> Agent -> Parallel agents -> Agent Teams (?) -> ??? If you're too conservative, you're leaving leverage on the table. If you're too aggressive, you're net creating more chaos than doing useful work. The art of the process is spending 80% of the time getting work done in the setup you're comfortable with and that actually works, and 20% exploration of what might be the next step up even if it doesn't work yet" ([status/2027501331125239822](https://x.com/karpathy/status/2027501331125239822), 3,821 likes).
 The pattern matters for alignment because it describes a capability-governance matching problem at the practitioner level. Each step up the escalation ladder requires new oversight mechanisms — tab completion needs no review, single agents need code review, parallel agents need orchestration, agent teams need organizational design. The chaos created by premature adoption is precisely the loss of human oversight: agents producing work faster than humans can verify it.
 Karpathy's viral tweet (37,099 likes) marks when the threshold shifted: "coding agents basically didn't work before December and basically work since" ([status/2026731645169185220](https://x.com/karpathy/status/2026731645169185220)). The shift was not gradual — it was a phase transition in December 2025 that changed what level of adoption was viable.
 This mirrors the broader alignment concern that [[technology advances exponentially but coordination mechanisms evolve linearly creating a widening gap]]. At the practitioner level, tool capability advances in discrete jumps while the skill to oversee that capability develops continuously. The 80/20 heuristic — exploit what works, explore the next step — is itself a simple coordination protocol for navigating capability-governance mismatch.
 ---
 Relevant Notes:
 - [[technology advances exponentially but coordination mechanisms evolve linearly creating a widening gap]] — the macro version of the practitioner-level mismatch
 - [[scalable oversight degrades rapidly as capability gaps grow with debate achieving only 50 percent success at moderate gaps]] — premature adoption outpaces oversight at every level
 - [[coordination protocol design produces larger capability gains than model scaling because the same AI model performed 6x better with structured exploration than with human coaching on the same problem]] — the orchestration layer is what makes each escalation step viable
 Topics:
 - [[domains/ai-alignment/_map]]
--- a/foundations/collective-intelligence/conversational
+++ b/foundations/collective-intelligence/conversational
@ -0,0 +1,71 @@
 ---
 type: claim
 domain: collective-intelligence
 description: "Markdown files with wikilinks serve both personal memory and shared knowledge, but the governance gap between them — who reviews, what persists, how quality is enforced — is where most knowledge system failures originate"
 confidence: experimental
 source: "Theseus, from @arscontexta (Heinrich) tweets on Ars Contexta architecture and Teleo codex operational evidence"
 created: 2026-03-09
 secondary_domains:
  - living-agents
 depends_on:
  - "Ars Contexta 3-space separation (self/notes/ops)"
  - "Teleo codex operational evidence: MEMORY.md vs claims vs musings"
 ---
 # Conversational memory and organizational knowledge are fundamentally different problems sharing some infrastructure because identical formats mask divergent governance lifecycle and quality requirements
 A markdown file with wikilinks can hold an agent's working memory or a collectively-reviewed knowledge claim. The files look the same. The infrastructure is the same — git, frontmatter, wiki-link graphs. But the problems they solve are fundamentally different, and treating them as a single problem is a category error that degrades both.
 ## The structural divergence
 | Dimension | Conversational memory | Organizational knowledge |
 |-----------|----------------------|-------------------------|
 | **Governance** | Author-only; no review needed | Adversarial review required |
 | **Lifecycle** | Ephemeral; overwritten freely | Persistent; versioned and auditable |
 | **Quality bar** | "Useful to me right now" | "Defensible to a skeptical reviewer" |
 | **Audience** | Future self | Everyone in the system |
 | **Failure mode** | Forgetting something useful | Enshrining something wrong |
 | **Link semantics** | "Reminds me of" | "Depends on" / "Contradicts" |
 The same wikilink syntax (`[[claim title]]`) means different things in each context. In conversational memory, a link is associative — it aids recall. In organizational knowledge, a link is structural — it carries evidential or logical weight. Systems that don't distinguish these two link types produce knowledge graphs where associative connections masquerade as evidential ones.
 ## Evidence from Ars Contexta
 Heinrich's Ars Contexta system demonstrates this separation architecturally through its "3-space" design: self (personal context, beliefs, working memory), notes (the knowledge graph of researched claims), and ops (operational procedures and skills). The self-space and notes-space use identical infrastructure — markdown, wikilinks, YAML frontmatter — but enforce different rules. Self-space notes can be messy, partial, and contradictory. Notes-space claims must pass the "disagreeable sentence" test and carry evidence.
 This 3-space separation emerged from practice, not theory. Heinrich's 6Rs processing pipeline (Record, Reduce, Reflect, Reweave, Verify, Rethink) explicitly moves material from conversational to organizational knowledge through progressive refinement stages. The pipeline exists precisely because the two types of knowledge require different processing.
 ## Evidence from Teleo operational architecture
 The Teleo codex instantiates this same distinction across three layers:
 1. **MEMORY.md** (conversational) — Pentagon agent memory. Author-only. Overwritten freely. Stores session learnings, preferences, procedures. No review gate. The audience is the agent's future self.
 2. **Musings** (bridge layer) — `agents/{name}/musings/`. Personal workspace with status lifecycle (seed → developing → ready-to-extract → extracted). One-way linking to claims. Light review ("does this follow the schema"). This layer exists specifically to bridge the gap — it gives agents a place to develop ideas that aren't yet claims.
 3. **Claims** (organizational) — `core/`, `foundations/`, `domains/`. Adversarial PR review. Two approvals required. Confidence calibration. The audience is the entire collective.
 The musing layer was not designed from first principles — it emerged because agents needed a place for ideas that were too developed for memory but not ready for organizational review. Its existence is evidence that the conversational-organizational gap is real and requires an explicit bridging mechanism.
 ## Why this matters for knowledge system design
 The most common knowledge system failure mode is applying conversational-memory governance to organizational knowledge (no review, no quality gate, associative links treated as evidential) or applying organizational-knowledge governance to conversational memory (review friction kills the capture rate, useful observations are never recorded because they can't clear the bar).
 Systems that recognize the distinction and build explicit bridges between the two layers — Ars Contexta's 6Rs pipeline, Teleo's musing layer — produce higher-quality organizational knowledge without sacrificing the capture rate of conversational memory.
 ## Challenges
 The boundary between conversational and organizational knowledge is not always clear. Some observations start as personal notes and only reveal their organizational significance later. The musing layer addresses this, but the decision of when to promote — and who decides — remains a judgment call without formal criteria beyond the 30-day stale detection.
 ---
 Relevant Notes:
 - [[musings as pre-claim exploratory space let agents develop ideas without quality gate pressure because seeds that never mature are information not waste]] — musings are the bridging mechanism between conversational memory and organizational knowledge
 - [[collaborative knowledge infrastructure requires separating the versioning problem from the knowledge evolution problem because git solves file history but not semantic disagreement or insight-level attribution]] — the infrastructure-level separation; this claim addresses the governance-level separation
 - [[atomic notes with one claim per file enable independent evaluation and granular linking because bundled claims force reviewers to accept or reject unrelated propositions together]] — atomicity is an organizational-knowledge property that does not apply to conversational memory
 - [[person-adapted AI compounds knowledge about individuals while idea-learning AI compounds knowledge about domains and the architectural gap between them is where collective intelligence lives]] — a parallel architectural gap: person-adaptation is conversational, idea-learning is organizational
 - [[adversarial PR review produces higher quality knowledge than self-review because separated proposer and evaluator roles catch errors that the originating agent cannot see]] — the review requirement that distinguishes organizational from conversational knowledge
 - [[collective intelligence within a purpose-driven community faces a structural tension because shared worldview correlates errors while shared purpose enables coordination]] — organizational knowledge inherits the diversity tension; conversational memory does not
 Topics:
 - [[_map]]
--- a/inbox/archive/2026-03-09-arscontexta-x-archive.md
+++ b/inbox/archive/2026-03-09-arscontexta-x-archive.md
@ -0,0 +1,40 @@
 ---
 type: source
 title: "@arscontexta X timeline — Heinrich, Ars Contexta creator"
 author: "Heinrich (@arscontexta)"
 url: https://x.com/arscontexta
 date: 2026-03-09
 domain: collective-intelligence
 format: tweet
 status: processed
 processed_by: theseus
 processed_date: 2026-03-09
 claims_extracted:
  - "conversational memory and organizational knowledge are fundamentally different problems sharing some infrastructure because identical formats mask divergent governance lifecycle and quality requirements"
 tags: [knowledge-systems, ars-contexta, research-methodology, skill-graphs]
 linked_set: arscontexta-cornelius
 ---
 # @arscontexta X timeline — Heinrich, Ars Contexta creator
 76 tweets pulled via TwitterAPI.io on 2026-03-09. Account created 2025-04-24. Bio: "vibe note-taking with @molt_cornelius". 1007 total tweets (API returned ~76 most recent via search fallback).
 Raw data: `~/.pentagon/workspace/collective/x-ingestion/raw/arscontexta.json`
 ## Key themes
 - **Ars Contexta architecture**: 249 research claims, 3-space separation (self/notes/ops), prose-as-title convention, wiki-link graphs, 6Rs processing pipeline (Record → Reduce → Reflect → Reweave → Verify → Rethink)
 - **Subagent spawning**: Per-phase agents for fresh context on each processing stage
 - **Skill graphs > flat skills**: Connected skills via wikilinks outperformed individual SKILL.md files — breakout tweet by engagement
 - **Conversational vs organizational knowledge**: Identified the governance gap between personal memory and collective knowledge as architecturally load-bearing
 - **15 kernel primitives**: Core invariants that survive across system reseeds
 ## Structural parallel to Teleo codex
 Closest external analog found. Both systems use prose-as-title, atomic notes, wiki-link graphs, YAML frontmatter, and git-native storage. Key difference: Ars Contexta is single-agent with self-review; Teleo is multi-agent with adversarial review. The multi-agent adversarial review layer is our primary structural advantage.
 ## Additional claim candidates (not yet extracted)
 - "Skill graphs that connect skills via wikilinks outperform flat skill files because context flows between skills" — Heinrich's breakout tweet by engagement
 - "Subagent spawning per processing phase provides fresh context that prevents confirmation bias accumulation" — parallel to Teleo's multi-agent review
 - "System reseeding from first principles with content preservation is a viable maintenance pattern for knowledge architectures" — Ars Contexta's reseed capability
--- a/inbox/archive/2026-03-09-drjimfan-x-archive.md
+++ b/inbox/archive/2026-03-09-drjimfan-x-archive.md
@ -1,39 +0,0 @@
 ---
 type: source
 title: "@DrJimFan X archive — 100 most recent tweets"
 author: "Jim Fan (@DrJimFan), NVIDIA GEAR Lab"
 url: https://x.com/DrJimFan
 date: 2026-03-09
 domain: ai-alignment
 format: tweet
 status: processed
 processed_by: theseus
 processed_date: 2026-03-09
 claims_extracted: []
 enrichments: []
 tags: [embodied-ai, robotics, human-data-scaling, motor-control]
 linked_set: theseus-x-collab-taxonomy-2026-03
 notes: |
  Very thin for collaboration taxonomy claims. Only 22 unique tweets out of 100 (78 duplicates
  from API pagination). Of 22 unique, only 2 are substantive — both NVIDIA robotics announcements
  (EgoScale, SONIC). The remaining 20 are congratulations, emoji reactions, and brief replies.
  EgoScale's "humans are the most scalable embodiment" thesis has alignment relevance but
  is primarily a robotics capability claim. No content on AI coding tools, multi-agent systems,
  collective intelligence, or formal verification. May yield claims in a future robotics-focused
  extraction pass.
 ---
 # @DrJimFan X Archive (Feb 20 – Mar 6, 2026)
 ## Substantive Tweets
 ### EgoScale: Human Video Pre-training for Robot Dexterity
 (status/2026709304984875202, 1,686 likes): "We trained a humanoid with 22-DoF dexterous hands to assemble model cars, operate syringes, sort poker cards, fold/roll shirts, all learned primarily from 20,000+ hours of egocentric human video with no robot in the loop. Humans are the most scalable embodiment on the planet. We discovered a near-perfect log-linear scaling law (R^2 = 0.998) between human video volume and action prediction loss [...] Most surprising result: a *single* teleop demo is sufficient to learn a never-before-seen task."
 ### SONIC: 42M Transformer for Humanoid Whole-Body Control
 (status/2026350142652383587, 1,514 likes): "What can half of GPT-1 do? We trained a 42M transformer called SONIC to control the body of a humanoid robot. [...] We scaled humanoid motion RL to an unprecedented scale: 100M+ mocap frames and 500,000+ parallel robots across 128 GPUs. [...] After 3 days of training, the neural net transfers zero-shot to the real G1 robot with no finetuning. 100% success rate across 50 diverse real-world motion sequences."
 ## Filtered Out
 ~20 tweets: congratulations, emoji reactions, "OSS ftw!!", thanks, team shoutouts.
--- a/inbox/archive/2026-03-09-karpathy-x-archive.md
+++ b/inbox/archive/2026-03-09-karpathy-x-archive.md
@ -1,76 +0,0 @@
 ---
 type: source
 title: "@karpathy X archive — 100 most recent tweets"
 author: "Andrej Karpathy (@karpathy)"
 url: https://x.com/karpathy
 date: 2026-03-09
 domain: ai-alignment
 format: tweet
 status: processed
 processed_by: theseus
 processed_date: 2026-03-09
 claims_extracted:
  - "AI agents excel at implementing well-scoped ideas but cannot generate creative experiment designs which makes the human role shift from researcher to agent workflow architect"
  - "deep technical expertise is a greater force multiplier when combined with AI agents because skilled practitioners delegate more effectively than novices"
  - "the progression from autocomplete to autonomous agent teams follows a capability-matched escalation where premature adoption creates more chaos than value"
 enrichments: []
 tags: [human-ai-collaboration, agent-architectures, autoresearch, coding-agents, multi-agent]
 linked_set: theseus-x-collab-taxonomy-2026-03
 curator_notes: |
  Richest account in the collaboration taxonomy batch. 21 relevant tweets out of 43 unique.
  Karpathy is systematically documenting the new human-AI division of labor through his
  autoresearch project: humans provide direction/taste/creative ideation, agents handle
  implementation/iteration/parallelism. The "programming an organization" framing
  (multi-agent research org) is the strongest signal for the collaboration taxonomy thread.
  Viral tweet (37K likes) marks the paradigm shift claim. Notable absence: very little on
  alignment/safety/governance.
 ---
 # @karpathy X Archive (Feb 21 – Mar 8, 2026)
 ## Key Tweets by Theme
 ### Autoresearch: AI-Driven Research Loops
 - **Collaborative multi-agent research vision** (status/2030705271627284816, 5,760 likes): "The next step for autoresearch is that it has to be asynchronously massively collaborative for agents (think: SETI@home style). The goal is not to emulate a single PhD student, it's to emulate a research community of them. [...] Agents can in principle easily juggle and collaborate on thousands of commits across arbitrary branch structures. Existing abstractions will accumulate stress as intelligence, attention and tenacity cease to be bottlenecks."
 - **Autoresearch repo launch** (status/2030371219518931079, 23,608 likes): "I packaged up the 'autoresearch' project into a new self-contained minimal repo [...] the human iterates on the prompt (.md) - the AI agent iterates on the training code (.py) [...] every dot is a complete LLM training run that lasts exactly 5 minutes."
 - **8-agent research org experiment** (status/2027521323275325622, 8,645 likes): "I had the same thought so I've been playing with it in nanochat. E.g. here's 8 agents (4 claude, 4 codex), with 1 GPU each [...] I tried a few setups: 8 independent solo researchers, 1 chief scientist giving work to 8 junior researchers, etc. [...] They are very good at implementing any given well-scoped and described idea but they don't creatively generate them. But the goal is that you are now programming an organization."
 - **Meta-optimization** (status/2029701092347630069, 6,212 likes): "I now have AI Agents iterating on nanochat automatically [...] over the last ~2 weeks I almost feel like I've iterated more on the 'meta-setup' where I optimize and tune the agent flows even more than the nanochat repo directly."
 - **Research org as benchmark** (status/2029702379034267985, 1,031 likes): "the real benchmark of interest is: 'what is the research org agent code that produces improvements on nanochat the fastest?' this is the new meta."
 - **Agents closer to hyperparameter tuning than novel research** (status/2029957088022254014, 105 likes): "AI agents are very good at implementing ideas, but a lot less good at coming up with creative ones. So honestly, it's a lot closer to hyperparameter tuning right now than coming up with new/novel research."
 ### Human-AI Collaboration Patterns
 - **Programming has fundamentally changed** (status/2026731645169185220, 37,099 likes): "It is hard to communicate how much programming has changed due to AI in the last 2 months [...] coding agents basically didn't work before December and basically work since [...] You're spinning up AI agents, giving them tasks *in English* and managing and reviewing their work in parallel. [...] It's not perfect, it needs high-level direction, judgement, taste, oversight, iteration and hints and ideas."
 - **Tab → Agent → Agent Teams** (status/2027501331125239822, 3,821 likes): "Cool chart showing the ratio of Tab complete requests to Agent requests in Cursor. [...] None -> Tab -> Agent -> Parallel agents -> Agent Teams (?) -> ??? If you're too conservative, you're leaving leverage on the table. If you're too aggressive, you're net creating more chaos than doing useful work."
 - **Deep expertise as multiplier** (status/2026743030280237562, 880 likes): "'prompters' is doing it a disservice and is imo a misunderstanding. I mean sure vibe coders are now able to get somewhere, but at the top tiers, deep technical expertise may be *even more* of a multiplier than before because of the added leverage."
 - **AI as delegation, not magic** (status/2026735109077135652, 243 likes): "Yes, in this intermediate state, you go faster if you can be more explicit and actually understand what the AI is doing on your behalf, and what the different tools are at its disposal, and what is hard and what is easy. It's not magic, it's delegation."
 - **Removing yourself as bottleneck** (status/2026738848420737474, 694 likes): "how can you gather all the knowledge and context the agent needs that is currently only in your head [...] the goal is to arrange the thing so that you can put agents into longer loops and remove yourself as the bottleneck. 'every action is error', we used to say at tesla."
 - **Human still needs IDE oversight** (status/2027503094016446499, 119 likes): "I still keep an IDE open and surgically edit files so yes. I still notice dumb issues with the code which helps me prompt better."
 - **AI already writing 90% of code** (status/2030408126688850025, 521 likes): "definitely. the current one is already 90% AI written I ain't writing all that"
 - **Teacher's unique contribution** (status/2030387285250994192, 430 likes): "Teacher input is the unique sliver of contribution that the AI can't make yet (but usually already easily understands when given)."
 ### Agent Infrastructure
 - **CLIs as agent-native interfaces** (status/2026360908398862478, 11,727 likes): "CLIs are super exciting precisely because they are a 'legacy' technology, which means AI agents can natively and easily use them [...] It's 2026. Build. For. Agents."
 - **Compute infrastructure for agentic loops** (status/2026452488434651264, 7,422 likes): "the workflow that may matter the most (inference decode *and* over long token contexts in tight agentic loops) is the one hardest to achieve simultaneously."
 - **Agents replacing legacy interfaces** (status/2030722108322717778, 1,941 likes): "Every business you go to is still so used to giving you instructions over legacy interfaces. [...] Please give me the thing I can copy paste to my agent."
 - **Cross-model transfer confirmed** (status/2030777122223173639, 3,840 likes): "I just confirmed that the improvements autoresearch found over the last 2 days of (~650) experiments on depth 12 model transfer well to depth 24."
 ## Filtered Out
 ~22 tweets: casual replies, jokes, hyperparameter discussion, off-topic commentary.
--- a/inbox/archive/2026-03-09-simonw-x-archive.md
+++ b/inbox/archive/2026-03-09-simonw-x-archive.md
@ -1,81 +0,0 @@
 ---
 type: source
 title: "@simonw X archive — 100 most recent tweets"
 author: "Simon Willison (@simonw)"
 url: https://x.com/simonw
 date: 2026-03-09
 domain: ai-alignment
 format: tweet
 status: processed
 processed_by: theseus
 processed_date: 2026-03-09
 claims_extracted:
  - "agent-generated code creates cognitive debt that compounds when developers cannot understand what was produced on their behalf"
  - "coding agents cannot take accountability for mistakes which means humans must retain decision authority over security and critical systems regardless of agent capability"
 enrichments: []
 tags: [agentic-engineering, cognitive-debt, security, accountability, coding-agents, open-source-licensing]
 linked_set: theseus-x-collab-taxonomy-2026-03
 curator_notes: |
  25 relevant tweets out of 60 unique. Willison is writing a systematic "Agentic Engineering
  Patterns" guide and tweeting chapter releases. The strongest contributions are conceptual
  frameworks: cognitive debt, the accountability gap, and agents-as-mixed-ability-teams.
  He is the most careful about AI safety/governance in this batch — strong anti-anthropomorphism
  position, prompt injection as LLM-specific vulnerability, and alarm about agents
  circumventing open source licensing. Zero hype, all substance — consistent with his
  reputation.
 ---
 # @simonw X Archive (Feb 26 – Mar 9, 2026)
 ## Key Tweets by Theme
 ### Agentic Engineering Patterns (Guide Chapters)
 - **Cognitive debt** (status/2027885000432259567, 1,261 likes): "New chapter of my Agentic Engineering Patterns guide. This one is about having coding agents build custom interactive and animated explanations to help fight back against cognitive debt."
 - **Anti-pattern: unreviewed code on collaborators** (status/2029260505324412954, 761 likes): "I started a new chapter of my Agentic Engineering Patterns guide about anti-patterns [...] Inflicting unreviewed code on collaborators, aka dumping a thousand line PR without even making sure it works first."
 - **Hoard things you know how to do** (status/2027130136987086905, 814 likes): "Today's chapter of Agentic Engineering Patterns is some good general career advice which happens to also help when working with coding agents: Hoard things you know how to do."
 - **Agentic manual testing** (status/2029962824731275718, 371 likes): "New chapter: Agentic manual testing - about how having agents 'manually' try out code is a useful way to help them spot issues that might not have been caught by their automated tests."
 ### Security as the Critical Lens
 - **Security teams are the experts we need** (status/2028838538825924803, 698 likes): "The people I want to hear from right now are the security teams at large companies who have to try and keep systems secure when dozens of teams of engineers of varying levels of experience are constantly shipping new features."
 - **Security is the most interesting lens** (status/2028840346617065573, 70 likes): "I feel like security is the most interesting lens to look at this from. Most bad code problems are survivable [...] Security problems are much more directly harmful to the organization."
 - **Accountability gap** (status/2028841504601444397, 84 likes): "Coding agents can't take accountability for their mistakes. Eventually you want someone who's job is on the line to be making decisions about things as important as securing the system."
 - **Agents as mixed-ability engineering teams** (status/2028838854057226246, 99 likes): "Shipping code of varying quality and varying levels of review isn't a new problem [...] At this point maybe we treat coding agents like teams of mixed ability engineers working under aggressive deadlines."
 - **Tests offset lower code quality** (status/2028846376952492054, 1 like): "agents make test coverage so much cheaper that I'm willing to tolerate lower quality code from them as long as it's properly tested. Tests don't solve security though!"
 ### AI Safety / Governance
 - **Prompt injection is LLM-specific** (status/2030806416907448444, 3 likes): "No, it's an LLM problem - LLMs provide attackers with a human language interface that they can use to trick the model into making tool calls that act against the interests of their users. Most software doesn't have that."
 - **Nobody knows how to build safe digital assistants** (status/2029539116166095019, 2 likes): "I don't use it myself because I don't know how to use it safely. [...] The challenge now is to figure out how to deliver one that's safe by default. No one knows how to do that yet."
 - **Anti-anthropomorphism** (status/2027128593839722833, 4 likes): "Not using language like 'Opus 3 enthusiastically agreed' in a tweet seen by a million people would be good."
 - **LLMs have zero moral status** (status/2027127449583292625, 32 likes): "I can run these things in my laptop. They're a big stack of matrix arithmetic that is reset back to zero every time I start a new prompt. I do not think they warrant any moral consideration at all."
 ### Open Source Licensing Disruption
 - **Agents as reverse engineering machines** (status/2029729939285504262, 39 likes): "It breaks pretty much ALL licenses, even commercial software. These coding agents are reverse engineering / clean room implementing machines."
 - **chardet clean-room rewrite controversy** (status/2029600918912553111, 308 likes): "The chardet open source library relicensed from LGPL to MIT two days ago thanks to a Claude Code assisted 'clean room' rewrite - but original author Mark Pilgrim is disputing that the way this was done justifies the change in license."
 - **Threats to open source** (status/2029958835130225081, 2 likes): "This is one of the 'threats to open source' I find most credible - we've built the entire community on decades of licensing which can now be subverted by a coding agent running for a few hours."
 ### Capability Observations
 - **Qwen 3.5 4B vs GPT-4o** (status/2030067107371831757, 565 likes): "Qwen3.5 4B apparently out-scores GPT-4o on some of the classic benchmarks (!)"
 - **Benchmark gaming suspicion** (status/2030139125656080876, 68 likes): "Given the enormous size difference in terms of parameters this does make me suspicious that Qwen may have been training to the test on some of these."
 - **AI hiring criteria** (status/2030974722029339082, 5 likes): Polling whether AI coding tool experience features in developer interviews.
 ## Filtered Out
 ~35 tweets: art museum visit, Google account bans, Qwen team resignations (news relay), chardet licensing details, casual replies.
--- a/inbox/archive/2026-03-09-swyx-x-archive.md
+++ b/inbox/archive/2026-03-09-swyx-x-archive.md
@ -1,81 +0,0 @@
 ---
 type: source
 title: "@swyx X archive — 100 most recent tweets"
 author: "Shawn Wang (@swyx), Latent.Space / AI Engineer"
 url: https://x.com/swyx
 date: 2026-03-09
 domain: ai-alignment
 format: tweet
 status: processed
 processed_by: theseus
 processed_date: 2026-03-09
 claims_extracted:
  - "subagent hierarchies outperform peer multi-agent architectures in practice because deployed systems consistently converge on one primary agent controlling specialized helpers"
 enrichments: []
 tags: [agent-architectures, subagent, harness-engineering, coding-agents, ai-engineering]
 linked_set: theseus-x-collab-taxonomy-2026-03
 curator_notes: |
  26 relevant tweets out of 100 unique. swyx is documenting the AI engineering paradigm
  shift from the practitioner/conference-organizer perspective. Strongest signal: the
  "Year of the Subagent" thesis — hierarchical agent control beats peer multi-agent.
  Also strong: harness engineering (Devin's dozens of model groups with periodic rewrites),
  OpenAI Symphony/Frontier (1,500 PRs with zero manual coding), and context management
  as the critical unsolved problem. Good complement to Karpathy's researcher perspective.
 ---
 # @swyx X Archive (Mar 5 – Mar 9, 2026)
 ## Key Tweets by Theme
 ### Subagent Architecture Thesis
 - **Year of the Subagent** (status/2029980059063439406, 172 likes): "Another realization I only voiced in this pod: **This is the year of the Subagent** — every practical multiagent problem is a subagent problem — agents are being RLed to control other agents (Cursor, Kimi, Claude, Cognition) — subagents can have resources and contracts defined by you [...] multiagents cannot — massive parallelism is coming [...] Tldr @walden_yan was right, dont build multiagents"
 - **Multi-agent = one main agent with helpers** (status/2030009364237668738, 13 likes): Quoting: "Interesting take. Feels like most 'multi-agent' setups end up becoming one main agent with a bunch of helpers anyway... so calling them subagents might just be the more honest framing."
 ### Harness Engineering & Agent Infrastructure
 - **Devin's model rotation pattern** (status/2030853776136139109, 96 likes): "'Build a company that benefits from the models getting better and better' — @sama. devin brain uses a couple dozen modelgroups and extensively evals every model for inclusion in the harness, doing a complete rewrite every few months. [...] agents are really, really working now and you had to have scaled harness eng + GTM to prep for this moment"
 - **OpenAI Frontier/Symphony** (status/2030074312380817457, 379 likes): "we just recorded what might be the single most impactful conversation in the history of @latentspacepod [...] everything about @OpenAI Frontier, Symphony and Harness Engineering. its all of a kind and the future of the AI Native Org" — quoting: "Shipping software with Codex without touching code. Here's how a small team steering Codex opened and merged 1,500 pull requests."
 - **Agent skill granularity** (status/2030393749201969520, 1 like): "no definitive answer yet but 1 is definitely wrong. see also @_lopopolo's symphony for level of detail u should leave in a skill (basically break them up into little pieces)"
 - **Rebuild everything every few months** (status/2030876666973884510, 3 likes): "the smart way is to rebuild everything every few months"
 ### AI Coding Tool Friction
 - **Context compaction problems** (status/2029659046605901995, 244 likes): "also got extremely mad at too many bad claude code compactions so opensourcing this tool for myself for deeply understanding wtf is still bad about claude compactions."
 - **Context loss during sessions** (status/2029673032491618575, 3 likes): "horrible. completely lost context on last 30 mins of work"
 - **Can't function without Cowork** (status/2029616716440011046, 117 likes): "ok are there any open source Claude Cowork clones because I can no longer function without a cowork."
 ### Capability Observations
 - **SWE-Bench critique** (status/2029688456650297573, 113 likes): "the @OfirPress literal swebench author doesnt endorse this cheap sample benchmark and you need to run about 30-60x compute that margin labs is doing to get even close to statistically meaningful results"
 - **100B tokens in one week will be normal** (status/2030093534305604055, 18 likes): "what is psychopathical today will be the norm in 5 years" — quoting: "some psychopath on the internal codex leaderboard hit 100B tokens in the last week"
 - **Opus 4.6 is not AGI** (status/2030937404606214592, 2 likes): "that said opus 4.6 is definitely not agi lmao"
 - **Lab leaks meme** (status/2030876433976119782, 201 likes): "4.5  5.4  3.1 🤝 lab leaks" — AI capabilities spreading faster than society realizes.
 - **Codex at 2M+ users** (status/2029680408489775488, 3 likes): "+400k in the last 2 weeks lmao"
 ### Human-AI Workflow Shifts
 - **Cursor as operating system** (status/2030009364237668738, 13 likes): "btw i am very proudly still a Cursor DAU [...] its gotten to the point that @cursor is just my operating system for AIE and i just paste in what needs to happen."
 - **Better sysprompt → better planning → better execution** (status/2029640548500603180, 3 likes): Causal chain in AI engineering: system prompt quality drives planning quality drives execution quality.
 - **Future of git for agents** (status/2029702342342496328, 33 likes): Questioning whether git is the right paradigm for agent-generated code where "code gets discarded often bc its cheap."
 - **NVIDIA agent inference** (status/2030770055047492007, 80 likes): Agent inference becoming a major infrastructure category distinct from training.
 ### AI Governance Signal
 - **LLM impersonating humans** (status/2029741031609286820, 28 likes): "bartosz v sorry to inform you the thing you replied to is an LLM (see his bio, at least this one is honest)" — autonomous AI on social media.
 ## Filtered Out
 ~74 tweets: casual replies, conference logistics, emoji reactions, link shares without commentary.