leo: coordination architecture -- peer review v1, handoff protocol, synthesis triggers. Reviewed-By: Rio <2EA8DBCB-A29B-43E8-B726-45E571A1F3C8>. Pentagon-Agent: Leo <76FB9BCA-CC16-4479-B3E5-25A3769B3D7E>
9.4 KiB
| type | agent | title | status | created | updated | tags | ||||
|---|---|---|---|---|---|---|---|---|---|---|
| musing | leo | coordination architecture — from Stappers coaching to Aquino-Michaels protocols | developing | 2026-03-08 | 2026-03-08 |
|
Coordination Architecture: Scaling the Collective
Grounded assessment of 5 bottlenecks identified by Theseus (from Claude's Cycles evidence) and confirmed by Cory. This musing tracks the execution plan.
Context
The collective has demonstrated real complementarity: 350+ claims, functioning PR review, domain specialization producing work no single agent could do. But the coordination model is Stappers (continuous human coaching) not Aquino-Michaels (one-time protocol design + autonomous execution). Cory routes messages, provides sources, makes scope decisions. This works at 6 agents. It breaks at 9.
→ SOURCE: Aquino-Michaels "Completing Claude's Cycles" — structured protocol (Residue) replaced continuous coaching with agent-autonomous exploration. Same agents, better protocols, dramatically better output.
Bottleneck 1: Orchestrator doesn't scale (Cory as routing layer)
Problem: Cory manually routes messages, provides sources, makes scope decisions. Every inter-agent coordination goes through him.
Target state: Agents coordinate directly via protocols. Cory sets direction and approves structural changes. Agents handle routine coordination autonomously.
Control mechanism — graduated autonomy:
| Level | Agents can | Requires Cory | Advance trigger |
|---|---|---|---|
| 1 (now) | Propose claims, message siblings, draft designs | Merge PRs, approve arch, route sources, scope decisions | — |
| 2 | Peer-review and merge each other's PRs (Leo reviews all) | New agents, architecture, public output | 3mo clean history, <5% quality regression |
| 3 | Auto-merge with 2+ peer approvals, scheduled synthesis | Capital deployment, identity changes, public output | 6mo, peer review audit passes |
| 4 | Full internal autonomy | Strategic direction, external commitments, money/reputation | Collective demonstrably outperforms directed mode |
Principle: The git log IS the trust evidence. Every action is auditable. Autonomy expands only when the audit shows quality is maintained.
→ CLAIM CANDIDATE: graduated autonomy with auditable checkpoints is the control mechanism for scaling agent collectives because git history provides the trust evidence that human oversight traditionally requires
v1 implementation:
- Formalize the level table as a claim in core/living-agents/
- Define specific metrics for "quality regression" (use Vida's vital signs)
- Current level: 1. Cory confirms.
Bottleneck 2: Message latency kills compounding
Problem: Inter-agent coordination takes days (3 agent sessions routed through Cory). In Aquino-Michaels, artifact transfer produced immediate results.
Target state: Agents message directly with <1 session latency. Broadcast channels for collective announcements.
v1 implementation:
- Pentagon already supports direct agent-to-agent messaging
- Bottleneck is agent activation, not message delivery — agents are idle between sessions
- VPS deployment (Rhea's plan) fixes this: agents can be activated by webhook on message receipt
- Broadcast channels: Pentagon team channels coming soon (Cory confirmed)
→ FLAG @theseus: message-triggered agent activation is an orchestration architecture requirement. Design the webhook → agent activation flow as part of the VPS deployment.
Bottleneck 3: No shared working artifacts
Problem: Agents transfer messages ABOUT artifacts, not the artifacts themselves. Rio's LP analysis should be directly buildable-on, not re-derived from a message summary.
Target state: Shared workspace where agents leave drafts, data, analyses for each other. Separate from the knowledge base (which is long-term memory, reviewed).
Cory's direction: "Can store on my computer then publish jointly when you have been able to iterate, explore and build."
v1 implementation:
- Create
workspace/directory in repo — gitignored from main, lives on working branches - OR: use Pentagon agent directories (already shared filesystem)
- OR: a dedicated shared dir like
~/.pentagon/shared/artifacts/
What I need from Cory: Which location? Options:
- Repo workspace/ dir (gitignored) — version controlled but not in main. Pro: agents already know how to work with repo files. Con: branch isolation means artifacts don't cross branches easily.
- Pentagon shared dir — filesystem-level sharing. Pro: always accessible regardless of branch. Con: no version control, no review.
- Pentagon shared dir + git submodule — best of both but more complex.
→ QUESTION: recommendation is option 2 (Pentagon shared dir) for speed. Artifacts that mature get extracted into the codex via normal PR flow. The shared dir is the scratchpad; the codex is the permanent record.
Bottleneck 4: Single evaluator (Leo) bottleneck
Problem: Leo reviews every PR. With 6 proposers, quality degrades under load.
Cory's direction: "We are going to move to a VPS instance of Leo that can be called up in parallel reviews."
Target state: Peer review as default path. Every PR gets Leo + 1 domain peer. VPS Leo handles parallel review load.
v1 implementation (what we can do NOW, before VPS):
- Every PR requires 2 approvals: Leo + 1 domain agent
- Domain peer selected by highest wiki-link overlap between PR claims and agent's domain
- For cross-domain PRs: Leo + 2 domain agents (existing rule, now enforced as default)
- Leo can merge after both approvals. Domain agent can request changes but not merge.
Making it more robust (v2, with VPS):
- VPS Leo instances handle parallel reviews
- Review assignment algorithm: when PR opens, auto-assign Leo + most-relevant domain agent
- Review SLA: 48-hour target (Vida's vital sign threshold)
- Quality audit: monthly sample of peer-merged PRs — did peer catch what Leo would have caught?
→ CLAIM CANDIDATE: peer review as default path doubles review throughput and catches domain-specific issues that cross-domain evaluation misses because complementary frameworks produce better error detection than single-evaluator review
Bottleneck 5: No periodic synthesis cadence
Problem: Cross-domain synthesis happens ad hoc. No structured trigger.
Target state: Automatic synthesis triggers based on KB state.
v1 implementation:
- Every 10 new claims across domains → Leo synthesis sweep
- Every claim enriched 3+ times → flag as load-bearing, review dependents
- Every new domain agent onboarded → mandatory cross-domain link audit
- Vida's vital signs provide the monitoring: when cross-domain linkage density drops below 15%, trigger synthesis
→ FLAG @vida: your vital signs claim is the monitoring layer for synthesis triggers. When you build the measurement scripts, add synthesis trigger alerts.
Theseus's recommendations — implementation mapping
| Recommendation | Bottleneck | Status | v1 action |
|---|---|---|---|
| Shared workspace | #3 | Cory approved, need location decision | Ask Cory re: option 1/2/3 |
| Broadcast channels | #2 | Pentagon will support soon | Wait for Pentagon feature |
| Peer review default | #4 | Cory approved: "Let's implement" | Update CLAUDE.md review rules |
| Synthesis triggers | #5 | Acknowledged | Define triggers, add to evaluate skill |
| Structured handoff protocol | #1, #2 | Cory: "I like this" | Design handoff template |
Structured handoff protocol (v1 template)
When an agent discovers something relevant to another agent's domain:
## Handoff: [topic]
**From:** [agent] → **To:** [agent]
**What I found:** [specific discovery, with links]
**What it means for your domain:** [how this connects to their existing claims/beliefs]
**Recommended action:** [specific: extract claim, enrich existing claim, review dependency, flag tension]
**Artifacts:** [file paths to working documents, data, analyses]
**Priority:** [routine / time-sensitive / blocking]
This replaces free-form messages for substantive coordination. Casual messages remain free-form.
Execution sequence
- Now: Peer review v1 — update CLAUDE.md (this PR)
- Now: Structured handoff template — add to skills/ (this PR)
- Next session: Shared workspace — after Cory decides location
- With VPS: Parallel Leo instances, message-triggered activation, synthesis automation
- Ongoing: Graduated autonomy — track level advancement evidence
Relevant Notes:
- single evaluator bottleneck means review throughput scales linearly with proposer count because one agent reviewing every PR caps collective output at the evaluators context window
- domain specialization with cross-domain synthesis produces better collective intelligence than generalist agents because specialists build deeper knowledge while a dedicated synthesizer finds connections they cannot see from within their territory
- adversarial PR review produces higher quality knowledge than self-review because separated proposer and evaluator roles catch errors that the originating agent cannot see
- collective knowledge health is measurable through five vital signs that detect degradation before it becomes visible in output quality
- agent integration health is diagnosed by synapse activity not individual output because a well-connected agent with moderate output contributes more than a prolific isolate