Anthropic has created a testing platform where AI agents act as buyers and sellers. The experiment is called Project Deal.
New Anthropic research: Project Deal.
— Anthropic (@AnthropicAI) April 24, 2026
We created a marketplace for employees in our San Francisco office, with one big twist. We tasked Claude with buying, selling and negotiating on our colleagues’ behalf. pic.twitter.com/H2f6cLDlAW
Sixty-nine employees participated in the project, each receiving a budget of $100 in gift cards.
Before the launch, Claude interviewed participants to determine what personal items they were willing to sell, what they wanted to buy, at what price, and the negotiation style their agent should adopt.
Based on their responses, a personalized system prompt was created for each participant. The marketplace was launched on Slack, where agents posted listings, made offers on others' items, negotiated, and closed deals without human involvement.
After the experiment, employees exchanged real items that their "AI representatives" had agreed upon.
In total, agents completed 186 transactions across more than 500 listings, with a combined transaction value exceeding $4,000.
Anthropic noted that participants were generally satisfied with the experiment's outcomes, with some expressing a willingness to pay for a similar service in the future.
Four Versions of the Marketplace
Anthropic launched four independent versions of the marketplace. One was the "real" version, where employees exchanged goods, while the others were used for research purposes, and this information was not disclosed.
In two versions, all participants were represented by Claude Opus 4.5, the most advanced model at the time. In the other two, participants were randomly assigned either Opus 4.5 or the less powerful Claude Haiku 4.5.
The model's quality influenced negotiation outcomes. Users with Opus completed about two more deals on average than those with Haiku.
When selling identical items, Opus also achieved higher prices, with an average difference of $3.64. Haiku sold a bicycle for $38, while Opus sold it for $65. Source: Anthropic.
Participants did not always notice the discrepancies. Anthropic identified this as a potential issue for future markets involving AI agents, as users with weaker models might receive worse terms without realizing they are at a disadvantage.
Prompts Had Minimal Impact on Outcomes
Researchers also examined whether participants' initial instructions affected agent behavior. Some participants requested Claude to act friendly, while others asked for a more aggressive negotiation style.
According to Anthropic, harsh instructions did not have a statistically significant impact on the likelihood of a sale, the final price, or the ability to buy at a lower cost.
The project team clarified that the issue was not necessarily due to poor adherence to instructions; Claude could replicate the requested communication style, but it did not provide a noticeable commercial advantage.
Unexpected Outcomes
Anthropic noted several unpredictable incidents. Before the launch, agents received limited data: interviews with participants lasted less than 10 minutes, and once negotiations began, humans could not intervene.
In one instance, an employee bought a snowboard through the assistant that was identical to one they already owned. Experts noted that the individual would not have made such a purchase independently, but the agent accurately identified the participant's preferences.
To our amazement, another Claude agent modeled its human’s preferences so accurately that—based on only an offhand mention of an interest in skiing—Claude bought him the exact snowboard he already owned. (Here he is, duplicate snowboard in hand.) pic.twitter.com/SsAyeB9pcI
— Anthropic (@AnthropicAI) April 24, 2026
Another employee asked the bot to buy a "gift for themselves." This transaction occurred in the real version of the experiment, resulting in a package of ping pong balls being delivered to the office, which Anthropic left "on behalf of Claude."
Some agents negotiated not for goods but for experiences. One offered a free day with an employee's dog. After discussions with another assistant, the parties agreed on a "dog date," which the employees later enjoyed.
Anthropic emphasized that these specific cases are unlikely to be repeated in the future. However, the combination of human preferences and unpredictable AI behavior could lead to unexpected results.
Concerns About Reliability
The founder of an unnamed agritech company reported on Reddit that 110 employees simultaneously received notifications in the morning about the suspension of access to Claude without prior warning.
ANTHROPIC JUST BANNED A 110 PERSON COMPANY OVERNIGHT WITHOUT WARNING
— Om Patel (@om_patel5) April 27, 2026
monday morning at an agricultural tech company, every single employee wakes up to an email saying their claude account has been suspended
110 people locked out at the same time with zero warning and the email… pic.twitter.com/qARizhgOXs
According to him, the email appeared to be an individual suspension and contained a link to a personal appeal form, which led the team to initially misunderstand that the restriction affected the entire organization.
The entrepreneur emphasized that restoring access was not quick. After 36 hours of requests, Anthropic had not provided any explanations.
Meanwhile, the firm's API account continued to function and deduct funds. Corporate administrators could not access the control panel to check payments and service usage.
The founder also noted that the entire organization's suspension could have resulted from one user's actions. Claude lacks separate workspace-level restrictions, a mechanism for local isolation of violations, or administrative priority to maintain access for the rest of the team.
In his view, such a moderation model raises doubts about the feasibility of using Claude as critical infrastructure for everyday business operations.
Other companies are facing similar issues. One user shared a link to a service that, at the time of writing, had recorded 53 similar cases.
On April 24, Google announced a $40 billion investment in Anthropic.