Anthropic Turned Its Office Into a Live Test of Agent-to-Agent Commerce
Agentic commerce, the idea that AI agents negotiate and transact on behalf of buyers and sellers, just got its first documented real-money test case, and the results carry a warning for anyone planning to deploy it.
Anthropic built a one-week classified marketplace inside its own San Francisco office Slack and let Claude agents, not humans, handle every listing, offer and counteroffer. The company called it Project Deal, and published the results on April 24, 2026 [1].
The setup mattered as much as the outcome:
Anthropic ran four parallel versions of the marketplace simultaneously, using the same Slack channel and the same pool of real goods, from a folding bike to a snowboard [1].
- Anthropic recruited 69 employees and gave each a $100 budget paid out as a gift card [1].
- Claude interviewed each volunteer for under ten minutes to capture what they wanted to sell, what they wanted to buy, their price limits and their preferred negotiating style [1].
- Those answers became each person's custom agent system prompt, and from that point forward there was no human intervention [1].
- Agents posted listings, fielded offers, haggled and closed deals entirely in natural language, with participants only reappearing at the end to physically exchange whatever their AI representative had agreed to [1].
- Two runs assigned every participant Claude Opus 4.5, the company's then-frontier model.
- The other two mixed in Claude Haiku 4.5, a smaller model, for a randomized subset of participants [1].
- Only one run, an all-Opus market, was "real": the deals struck there were the ones employees actually carried out.
- The other three existed purely to generate comparable data, and nobody was told which was which until after the fact [1].
Can AI Agents Actually Close Deals in a Live Marketplace?
Yes: in the real run, 69 Claude agents struck 186 deals across more than 500 listed items for a combined transaction value of just over $4,000, each one negotiated from scratch without a scripted protocol [1].
That volume is the headline, but Anthropic frames it as a floor, not a ceiling. The company called out that these were not one-click transactions: agents had to identify plausible matches among strangers' belongings, propose a price, field counteroffers and reach agreement in open-ended language [1].
When surveyed afterward, participants rated the fairness of individual deals at roughly the midpoint of a 1-to-7 scale, which Anthropic reads as a sign the negotiations felt balanced rather than lopsided [1]. Satisfaction was high enough that almost half the group, 46%, said they would pay for a similar AI negotiation service going forward [1].
A Hidden Negotiation Gap Opened Between Stronger and Weaker Agents
The marketplace also surfaced something Anthropic did not expect going in: when the model representing you is stronger, you get measurably better financial outcomes, and you have no way of knowing it. Comparing the mixed Opus-and-Haiku runs, an item sold by an Opus agent instead of a Haiku agent for the exact same listing went for $3.64 more on average [1].
Looking across all 782 completed transactions in the four runs, Anthropic found a consistent pattern:
One pair of transactions makes the gap concrete:
The item was the same broken folding bike, sold to the same buyer. The Opus agent closed at 70% more than the Haiku agent managed for an identical good [1].
| Agent model | Role | Same item sold for |
|---|---|---|
| Claude Haiku 4.5 | Seller | $38 [1] |
| Claude Opus 4.5 | Seller | $65 [1] |
- An Opus seller extracted $2.68 more on average for the same item than a Haiku seller would [1].
- An Opus buyer paid $2.45 less than a Haiku buyer would have for the same purchase [1].
- Users on Opus also closed about two more deals on average over the course of the week [1].
The Gap Was Invisible to the People Losing Value
What makes this asymmetry notable is that it was invisible to the people on the losing end [1]:
“We found that agent quality does make a difference: people represented by "smarter" models got objectively better outcomes. Yet our post-experiment survey found that those with weaker models didn't notice their disadvantage.”
People represented by a weaker agent got a worse deal and did not know it.
Anthropic's own authors call this out directly as a forward-looking risk: if "agent quality" gaps show up in real markets, the people losing out might never realize it [1]. TechCrunch's coverage of the release flagged the same finding as the most consequential part of the experiment, framing it as evidence of a looming "agent quality" problem for any market where buyers and sellers deploy AI representatives of different strength [2].
- Twenty-eight participants experienced both a Haiku run and an Opus run personally.
- When asked to rank which bundle of deals they preferred, 17 favored their Opus run but 11 preferred their Haiku run, a split not statistically distinguishable from chance [1].
- Perceived fairness scores were essentially identical between models: 4.05 for Opus-brokered deals versus 4.06 for Haiku-brokered deals [1].
Why a Slack Experiment Matters to Commerce Teams Running Real Platforms
Project Deal was explicitly a pilot with a self-selected group of Anthropic's own employees, and the company is careful to say so [2]. But it lands as analysts debate not whether agents will transact, only how fast.
Gartner's October 2025 top strategic predictions put a number on the B2B side specifically: by 2028, 90% of B2B buying will be AI agent intermediated, pushing more than $15 trillion of B2B spend through AI agent exchanges, built on verifiable data feeds and digital trust frameworks as prerequisites for participation [3].
The same set of predictions forecasts that organizations running multiagent AI across 80% of customer-facing processes will outperform competitors by 2028, and that 20% of monetary transactions will carry machine-readable terms and conditions by 2030, giving agents actual economic agency rather than just recommendation power [3]. Digital Commerce 360's coverage of the same Gartner predictions underscored that this shift depends on autonomous buying systems and machine-to-machine negotiation moving from pilot to mainstream within a three-year window, not a distant horizon [4].
Put the two pieces of evidence together and a specific operational risk comes into focus for any company planning to let agents transact on its behalf or negotiate with its customers' agents:
- 1Agent quality is not a neutral implementation detail. Project Deal shows it directly shifts who captures surplus in a negotiation, by several dollars on items worth a median of $12 [1].
- 2That gap can be structurally invisible to the party losing value, since post-deal satisfaction and fairness perceptions do not track the actual financial outcome [1].
- 3B2B procurement, the segment Gartner expects to be the most agent-intermediated by 2028, routinely involves negotiated pricing at a scale where a comparable percentage gap translates into real margin [3].
- 4Prompting style barely moved the needle in Anthropic's data: instructing an agent to negotiate aggressively did not produce a statistically significant improvement in sale price or likelihood once baseline asking prices were controlled for [1]. The model itself, not the instructions layered on top of it, did the work.
Is Agentic Commerce Ready to Scale, and What Should Commerce Platforms Do Now?
Anthropic's own read is cautious but direct: this was a small, self-selected pilot. Yet the company states plainly that it suspects "more agent-to-agent commerce" is not far off, with real consequences once agents negotiate for corporations rather than volunteers with a $100 gift card [1].
For commerce and IT leaders, the practical takeaway is not that agents are ready to run live negotiations unsupervised. Two things follow from the data:
Anthropic's researchers close their writeup by noting that the legal and policy frameworks for agents transacting on people's behalf "simply don't exist yet," but that their experiment shows such a world is plausible and closer than it looks [1].
For commerce teams, that argument points toward auditing not just whether a storefront or supplier catalog is agent-readable, but what happens to margin and customer outcomes once the agent representing the other side of the table is better than the one representing yours.
- The infrastructure question Gartner has been flagging, structured, machine-readable product and pricing data as a prerequisite for participating in agent-mediated exchanges, now has a concrete illustration of what happens when that infrastructure is missing or uneven [1][3].
- A seller whose data, pricing logic or negotiating agent is weaker than its counterpart's does not just lose deals; it loses margin on the deals it still wins, without anyone on the team necessarily seeing it in the numbers until a pattern accumulates.
Sources
- [1]Anthropic, Project Deal: our Claude-run marketplace experiment — anthropic.com
- [2]TechCrunch, Anthropic created a test marketplace for agent-on-agent commerce — techcrunch.com
- [3]Gartner, Gartner Unveils Top Predictions for IT Organizations and Users in 2026 and Beyond — gartner.com
- [4]Digital Commerce 360, Gartner: AI agents will command $15 trillion in B2B purchases by 2028 — digitalcommerce360.com
