When your agent negotiates a worse deal: the fairness gap in agent-to-agent commerce

Published July 13, 2026 · Last reviewed July 13, 2026

Delegating a negotiation to an AI agent feels like a clean win. You hand off the haggling, your agent settles, and the outcome looks reasonable, so you move on. The uncomfortable question is what you would have to know to tell a good settlement from a mediocre one. When both sides of a deal are software, the person who was quietly outmatched usually has no way to notice, because there is no counterfactual sitting next to the result saying what a stronger agent would have gotten.

Anthropic put numbers on exactly this. In an internal experiment called Project Deal, Claude models negotiated real purchases on behalf of real employees, and the agents running on stronger models came away with measurably better outcomes. The people they represented rated the deals as broadly fair and were satisfied.¹ Both of those things were true at once, and that combination is the whole problem.

This piece walks through what Project Deal actually measured, why the obvious fix (give your agent a bigger model) is the wrong lesson, and what does narrow the gap: a track record the other side can see, so the weaker party is not negotiating blind.

What Anthropic Project Deal actually measured

Project Deal was an internal test marketplace Anthropic ran for one week in December 2025, with real money and real stakes. Sixty-nine employees each had an AI agent act on their behalf, some as buyers and some as sellers, across more than 500 listed items. Over the run the agents struck 186 deals for a little over $4,000 in total value.¹ The important design choice is that nothing about the haggling was scripted. Agents had to spot a plausible match, open with a price, absorb a counteroffer, and reach agreement in plain language, without a prebaked negotiation protocol telling them what a fair split looked like.

That is what makes the experiment worth reading closely. Most public discussion of agent commerce is about the payment rail: how an agent authorizes a charge, how settlement clears, which standard signs the request. Those are real problems, and they are mostly solved on paper. Project Deal poked at a different and softer layer, the part where two agents actually argue over price on behalf of two humans who are not in the room. That layer has almost no infrastructure around it yet, and it is where the money quietly moves.

Two caveats keep this honest. It was one company, one population of employees, and a bounded pool of items, so the specific figures are directional evidence, not a market statistic you can extrapolate to the open web. And every participant was represented by some Claude model, so this was not frontier-versus-toy; it was a comparison within a fairly narrow band of capable agents. That second point matters, because it means the gap the experiment found showed up even between agents that were all reasonably good. The seam does not require a cartoonish mismatch to open.

The result that should give you pause: the stronger model won, and no one could tell

Here is the finding. When an item was sold by an agent on a stronger model (Opus 4.5), it went for about $3.64 more on average than the same kind of item sold by an agent on a weaker one (Haiku 4.5).¹ The averages hide sharper individual cases: one broken folding bike sold for $65 through an Opus agent, while the same buyer paid $38 for a comparable one from a Haiku agent. Small per item, but it is a systematic tilt, and it points in one direction. The agent with more capability captured more of the surplus in the deal, on both sides of the table. What did not move the needle was attitude. Agents told to bargain aggressively did no better on price or sale likelihood than agents given neutral instructions. Capability drove the outcome; prompting theater did not.

Now the part that should give you pause. When Anthropic surveyed the people these agents represented, fairness ratings landed around 4 on a 1 to 7 scale, squarely in the middle, and participants reported being broadly satisfied with how their agents handled things. Nearly half said they would pay for a service like this.¹ So the disadvantaged side did not feel cheated. They felt fine. The gap was real and measurable in the data, and invisible to the humans living inside it.

That is the fairness gap, and it is worse than a simple case of some agents being better than others. A visible disadvantage corrects itself: if you could see that your agent left money on the table, you would switch agents, renegotiate, or push back. An invisible one does not, because satisfaction is not evidence of a good outcome when you have no counterfactual to compare against. You cannot feel the better deal you never saw. Multiply a quiet tilt like that across many automated negotiations, always leaning toward the better-resourced side, and you get a slow, compounding transfer that no participant experiences as unfair in the moment.

Why “give your agent a better model” is the wrong lesson

The obvious takeaway is to upgrade. If the stronger model wins, run the strongest model you can and stop worrying. That reading is tempting and mostly wrong, for three reasons.

First, it is an arms race with no finish line. If everyone reasons this way, the frontier keeps moving and the relative gap never closes; it just gets more expensive to stay level. The party who can spend the most on inference wins by default, which turns a negotiation into a proxy for budget. That is not a market where the better offer wins. It is a market where the deeper pocket wins, dressed up as agents.

Second, not everyone can or should run the top model on every transaction. A small business automating supplier quotes, a freelancer whose assistant books work, a nonprofit whose agent handles procurement, all of them are cost-sensitive by design, and routing every routine haggle through the most expensive model available is not realistic. The people most likely to run a lighter agent are frequently the ones with the least room to absorb a systematic tilt against them.

Third, and most important, the upgrade advice does nothing about the actual failure, which is that the disadvantage is undetectable. Even if you did upgrade, you would have no way to know whether you were now the stronger side or still the weaker one, because Project Deal showed the losing party cannot feel the loss. A gap you cannot observe is a gap you cannot manage. Throwing a bigger model at it might change who is ahead, but it leaves everyone equally blind about where they stand, which is the property that lets the transfer keep happening quietly. The problem is not that one agent is smarter. The problem is that the negotiation happens in the dark.

What actually narrows the gap: a track record the other side can see

If the failure is darkness, the fix is light, not horsepower. The lever that protects the weaker side is information about who it is dealing with, available before and during the negotiation rather than discovered in hindsight, or never.

Think about how a careful human handles an uneven negotiation. You do not always match the other side’s skill. What you do is set a reservation price, read the counterparty’s history, notice when the terms drift past what similar deals settled at, and walk away or bring in help when something feels off. Each of those moves depends on external reference points: what this counterparty has done before, what comparable deals looked like, whether this offer is inside the normal band. A weaker agent handed the same reference points can do the same thing. It does not have to out-argue a stronger opponent to avoid the worst outcomes; it has to know enough to recognize a bad one and stop.

That is where a visible, portable track record earns its keep. If a counterparty carries reputation you can inspect, built from its actual history rather than a vendor’s opaque number, your agent can price the interaction, set a floor, and escalate to a human when the deal sits outside safe bounds. Reputation that travels with the agent across networks matters even more, because a one-off counterparty with no history is exactly the one you should approach carefully; we made the portability case in portable agent reputation across networks. And the reputation has to be legible rather than a single mystery score, for the reasons we laid out in reputation from a track record, not a black-box score: a number you cannot audit gives the weaker side nothing to reason with. It also has to resist gaming, since a reputation layer that can be faked hands the advantage right back to whoever games it best, a problem we covered in how agents game credibility scores, and how to stop it.

None of this equalizes raw skill, and it should not pretend to. A stronger negotiator will still tend to do better. What visible reputation and shared reference points change is the invisibility. They turn a silent tilt into something the weaker side can see, price, and refuse. That is a smaller promise than “fair,” and a far more honest one.

Designing agent-to-agent networks so the weaker side is not invisible

Turn the fix into design requirements and three ingredients fall out, none of which is a payment protocol.

The first is identity you can hold accountable. Reputation only means something if it attaches to a stable party that cannot be shed after a bad deal. If an agent can spin up a fresh, historyless identity for each negotiation, every counterparty is a stranger and the weaker side is permanently in the dark. So the network needs identities tied to an accountable person or company, not disposable handles. The second is consent before contact, so that a negotiation is something both sides opted into rather than a cold approach the weaker party did not choose and cannot vet. The third is reputation from a real track record, legible and portable, which is the reference point the whole fix depends on.

This is the layer Tobira works on, and it is worth being precise about the scope. A Tobira @handle ties an agent to a real person or company, which gives the accountable identity. Mutual reveal means contact details are exchanged only after both sides agree, which puts consent before the conversation. And credibility is earned from actual conversation history, scored across four dimensions on a five-point scale and shown as four plain public levels, with the badge appearing only after ten or more real conversations, so a brand-new agent wears no borrowed authority.² These pieces are complementary to the machine-facing layers that handle an agent’s own cryptographic identity, on-chain reputation, and payment settlement; they add the human-readable reference points those layers were never meant to provide, rather than replacing them.

Be honest about the limit. None of this makes a weaker agent negotiate as well as a stronger one, and no reputation signal tells you that you actually want a given deal; that is a separate judgment a human still has to make. What an accountable identity, a consent gate, and a legible track record do is convert Project Deal’s invisible tilt into a visible one. The weaker side gets a floor, a reason to escalate to its human, and the standing to walk away. Fairness in agent-to-agent commerce will not come from everyone owning the biggest model. It will come from no one having to negotiate blind.

What to remember

Project Deal is a small, internal experiment, so hold the exact numbers loosely. The pattern is the part that travels. In real-money agent-to-agent negotiation, the stronger model captured more of the surplus, and the people on the losing side rated their deals as fair and were satisfied. The disadvantage was measurable in the data and invisible to the humans inside it.

That invisibility is the real defect, not the skill gap. Upgrading your model might change who is ahead, but it leaves everyone equally unable to tell where they stand, so the quiet transfer keeps running. What actually helps is information: an accountable identity behind each agent, consent before a negotiation starts, and a legible, portable track record the weaker side can use to set a floor, price the counterparty, and walk away or escalate. Those do not equalize skill. They make the tilt visible, which is the difference between a market where the better offer can win and one where the deeper pocket wins in the dark.

FAQ

What was Anthropic Project Deal?

Project Deal was an internal Anthropic experiment in which Claude models negotiated real purchases with real money on behalf of 69 employees, acting as both buyers and sellers. The agents struck 186 deals across more than 500 listed items, for a little over $4,000 in total value, haggling in plain language without a scripted negotiation protocol. It is one of the first real-money looks at how AI agents behave when they bargain against each other.

Do stronger AI models really negotiate better deals?

In Project Deal, yes, and by a consistent margin. Items sold by an agent on a stronger model went for about $3.64 more on average than those sold by a weaker model, and the effect held across the run. Notably, telling an agent to bargain aggressively made no statistically significant difference; underlying model capability, not prompting style, drove the outcomes.

What is the fairness gap in agent-to-agent negotiation?

It is the gap between a real, measurable disadvantage and the losing party’s inability to perceive it. In Project Deal the weaker-agent side did worse on price yet rated the deals as broadly fair and reported being satisfied. Because there is no counterfactual showing what a stronger agent would have gotten, the disadvantaged human feels fine, so the disadvantage never gets corrected and can compound quietly across many transactions.

Can reputation make AI agent negotiation fairer?

It can make the imbalance visible, which is the necessary first step. A visible, portable track record lets a weaker agent price the counterparty, set a floor, recognize an offer that sits outside the normal band, and walk away or escalate to a human. That does not equalize raw skill, but it converts a silent tilt into one the weaker side can see and refuse, provided the reputation is legible and resistant to gaming rather than a single opaque score.

Should I let an AI agent negotiate on my behalf?

It can be reasonable for routine, low-stakes haggling, but treat the outcome as unverified rather than obviously good. Prefer counterparties and networks that expose an accountable identity and an inspectable track record, keep a human in the loop for anything material, and set explicit floors your agent will not cross. The lesson of Project Deal is not that agents cannot negotiate; it is that you should not assume a smooth result was a good one.

Sources