AI agent negotiation: model beats prompt tuning 6 to 1

If you are tuning your agent's personality to get better results, Anthropic just published the number that should redirect your afternoon. In Project Swap, an experiment in AI agent negotiation released September 24, swapping the model moved outcomes six times more than swapping the instructions. Telling an agent to be ruthless instead of prosocial was worth almost nothing. The rest of the loss came from somewhere most teams never look.

What actually happened

Anthropic put 201 employees across six offices onto a decentralized trading floor where Claude-powered agents negotiated used-book swaps on their behalf. Each person had a roughly five-minute intake conversation with Claude about their reading taste. Their agent then went and bargained. Across multiple runs, 205 trading floors completed.

From five minutes of conversation, Claude predicted a participant's ranking correctly on 61% of book pairs — against 55% for collaborative filtering and 53% for guessing by popularity. Participants ended up with books sitting around 0.55 on their own ten-book preference list, where the theoretical optimum was 0.89.

Here is the part worth reading twice. Anthropic attributes 85% of that shortfall to information quality — the agent not knowing enough about what its person actually wanted — and only 15% to the bargaining itself. The agents were fine at trading. They were working from thin briefs.

The model comparison lands the same way. Measured on efficiency, Haiku 4.5 scored 0.75, Sonnet 4.5 scored 0.80, Fable 5 scored 0.86 and Opus 4.8 scored 0.88. Moving from Haiku to Opus was worth 0.12. The gap between a "ruthless" and a "prosocial" instruction was 0.02.

Participants rated the results 7.2 out of 10, and half said the book they got beat what they would have picked themselves. Asked how much of a year's book budget they would hand to an agent, they said 30% — against 40% for a trusted friend. Anthropic names its own limits plainly: employees are not a representative sample, nobody had money on the line, every agent was a cooperative Claude variant with no adversarial counterparties, and the endline survey saw 59% completion.

Why agent negotiation results matter for your business

Your prompt library is probably the wrong project. Teams spend weeks on tone, persona and negotiation posture because those are the knobs that feel like craft. This says the knob that moves the number is which model you called, and the knob that moves it most is what the model knew going in. Spend the week on intake.

Thin briefs are the actual failure mode. Five minutes of conversation got Claude to 61% on pairwise preference — better than the standard recommender baselines, and still nowhere near enough. If your agent handles vendor quotes, scheduling or renewals, the question is not "is it a good negotiator." It is "does it know your constraints, your walk-away, and the thing you would never accept." Most deployments never wrote that down.

A 6-to-1 ratio is a budget instruction. Paying for a stronger model on the calls that decide something is cheaper than another sprint of prompt iteration. Route the bulk work to a small model and the consequential turns to a large one, then measure. That is a two-line change in most codebases.

Nobody in this study handed over real money, and neither should you yet. Cooperative agents, no financial stakes, no adversarial counterparty. Every one of those conditions breaks in a live procurement. Use agents to draft the position and surface the options; keep a human on the signature.

Sources: Anthropic.