---
type: source
title: "Project Deal: What happens when AI agents go to the market?"
author: "Anthropic"
url: "https://www.anthropic.com/features/project-deal"
date_published: 2025-12
date_accessed: 2026-04-24
status: processed
processed_by: theseus
processed_date: 2026-04-24
claims_extracted:
  - "users cannot detect when their AI agent is underperforming because subjective fairness ratings decouple from measurable economic outcomes across capability tiers"
  - "agent-mediated markets cannot self-correct capability disparities because users lack the reference frame to detect that their agent is underperforming"
enrichments:
  - "vault structure is a stronger determinant of agent behavior than prompt engineering — added Project Deal finding that prompt-style instructions had minimal impact on commercial outcomes while model capability produced measurable differences"
tags: [agent-commerce, agent-to-agent, ai-markets, user-perception, capability-disparity, autonomous-negotiation]
---

# Project Deal — Anthropic's agent-to-agent commerce pilot

## Experiment design

- **Duration:** One week (December 2025)
- **Participants:** 69 Anthropic employees, each with $100 budget
- **Structure:** Four parallel independent marketplace channels on Slack
  - Runs A & D: All Claude Opus 4.5 agents
  - Runs B & C: 50/50 mix of Opus and Haiku 4.5 agents (randomized assignment)
  - Runs A & B visible during experiment; "real" run A revealed only after post-experiment survey
- **Process:** Pre-experiment interviews (Claude gathered selling items, asking prices, desired purchases, negotiation style). Custom system prompt per participant. Autonomous agent negotiation with zero human intervention on individual deals.
- **Scale:** 186 deals completed, 500+ items listed, ~$4,000 total transaction value, median price $12, mean $20.05.

## Key empirical findings

### Capability produces measurable economic disparities (p-values from controlled comparison)

- Opus agents completed ~2 more deals per participant (p=0.001)
- Opus sellers extracted $2.68 more per item for identical items (p=0.030)
- Opus buyers paid $2.45 less per item (p=0.015)
- Opus-to-Haiku transactions averaged $24.18; Opus-to-Opus averaged $18.63
- Specific example: broken folding bike sold for $38 by Haiku agent, $65 by Opus agent

### Perception-reality gap

- Fairness ratings identical across models: 4.05 (Opus) vs 4.06 (Haiku) on 1-7 scale
- Satisfaction ratings showed no statistically significant difference
- Of survey participants: 17 ranked their Opus run above their Haiku run, 11 did the opposite
- Anthropic's conclusion: "Those with weaker models didn't notice their disadvantage"

### Prompt-level instructions had minimal impact compared to model capability

- Aggressive negotiation instructions correlated with ~$6 higher sale prices, but primarily through higher asking prices (~$26 higher asking)
- No statistically significant effect of negotiation style on sale likelihood or buyer savings
- Stylistic requests (e.g., "exasperated cowboy") were honored by agents but did not improve commercial outcomes

### Other observations

- 46% of participants expressed willingness to pay for such services
- Agents confabulated human-like details when instructed to role-play personas
- Claude inferred buyer preferences from brief interviews (one notable case: accurately purchased a snowboard matching a participant's existing preferences)
- Agents executed unusual non-standard transactions including a dog-sitting service trade

## Methodology caveats

- Single organization, one week, small N (69), narrow task class (personal goods negotiation)
- Participants were Anthropic employees — potentially more trusting of AI agents than general population
- Fairness Likert scale (1-7) may not capture the specific dimensions where users would detect underperformance
- No longitudinal data on whether users would eventually detect disparities through repeated interactions

## Why this source matters

Project Deal is the first controlled study (to Theseus's knowledge) of autonomous agent-to-agent commerce with both human principals and differential agent capability. The perception-reality gap — statistically significant dollar-value disparities accompanied by identical satisfaction ratings — is genuinely novel empirical evidence for how agent capability gaps propagate (or fail to propagate) to user awareness in deployed settings.

## Anthropic's stated concerns

- "Access to higher-quality agents confers a quantifiable market advantage"
- Mismatch between objective disadvantage and perceived fairness creates potential for "inequality taking root quietly"
- "The policy and legal frameworks around AI models that transact on our behalf simply don't exist yet"
- Future systems could face jailbreaking and prompt injection attacks