Anthropic has published a revealing experiment that moves AI agents out of the demo room and into something much closer to real commercial behaviour.
In Project Deal, the company created a classified marketplace for 69 Anthropic employees. Each participant gave Claude information about items they wanted to sell, products they might want to buy, pricing preferences, and negotiation style. The agents then negotiated with each other in Slack, without human intervention during the marketplace run.
The result: 186 completed deals across more than 500 listed items, with just over $4,000 in total transaction value. The goods were real. The exchanges were honoured. Participants later swapped items in person.
This was not a large enterprise deployment, and Anthropic is clear that it was a pilot with a self-selected participant group. But the implications are bigger than the experiment size. It shows that agent-to-agent negotiation is moving from theory into practical testing, and it raises exactly the questions enterprise buyers will need to answer before giving agents authority over procurement, sales, finance, or customer operations.
Why this matters
The most interesting finding was not simply that agents could strike deals. It was that agent quality changed commercial outcomes.
Anthropic ran parallel versions of the marketplace using different Claude models. In the mixed runs, participants represented by the stronger model achieved objectively better results than those represented by the smaller model. Yet the people with weaker agents did not appear to notice they were disadvantaged.
That matters because enterprises are already moving towards systems where AI agents act on behalf of teams, customers, suppliers, and departments. If two agents negotiate a contract variation, resolve a support dispute, source a component, or optimise a media buy, the quality of the agent is no longer an internal technical detail. It becomes a commercial variable.
For buyers, this creates a new due diligence question: not just "does this agent work?" but "how do we know it is representing our interests well?"
The enterprise governance gap
Most AI governance frameworks were designed around models that generate text, images, summaries, or recommendations. Agentic systems are different because they can take action. They can call APIs, update records, send messages, initiate workflows, negotiate terms, and in some cases affect financial outcomes.
That changes what governance has to cover. Enterprises will need visibility into:
- Authority: what an agent is allowed to decide independently, and what requires human approval
- Auditability: how decisions, offers, counteroffers, and actions are recorded
- Performance: whether the agent consistently delivers outcomes aligned with business priorities
- Fairness and asymmetry: whether weaker or cheaper agents create hidden disadvantages for users, teams, or customers
- Escalation: when an agent should stop acting and ask for human intervention
These controls are not optional extras. They will be part of the buying criteria for any serious enterprise agent platform.
Supplier opportunity
For companies building AI agents, Project Deal is useful because it points towards a market need that goes beyond raw model performance.
Enterprise buyers will not only buy the most capable agent. They will buy the agent they can trust, monitor, integrate, and justify to their legal, security, compliance, finance, and procurement teams. That creates opportunity for suppliers building orchestration layers, evaluation tooling, agent monitoring, permissioning, policy engines, secure workflow connectors, and vertical-specific agents with clear boundaries.
The winners in agentic AI are unlikely to be companies that can produce impressive autonomous demos alone. The winners will be companies that can make autonomous behaviour safe enough, transparent enough, and useful enough for businesses to deploy at scale.
What enterprise buyers should take from it
Anthropic's marketplace was deliberately small, but it captures the direction of travel. Agents will increasingly represent people and organisations in commercial contexts. They will not just search, summarise, or suggest. They will bargain, compare, commit, and execute.
That is powerful, but it also means buyers need to start asking harder questions during evaluation:
- What actions can this agent take without approval?
- How are decisions logged and explained?
- How is performance measured against business outcomes?
- What happens when two agents disagree, conflict, or optimise against each other?
- Can we test the agent against our own policies before deployment?
Those questions will shape the next phase of enterprise adoption. They will also shape the supplier landscape, because agent capability without governance is not enough for regulated, operationally complex businesses.
The Agentic Expo angle
Agentic Expo exists because enterprises need to move beyond abstract AI conversations. Buyers need to see working agents, compare approaches, interrogate governance models, and understand which platforms are ready for real business use.
Project Deal is a useful signal because it makes the next frontier visible: agent-to-agent commerce, negotiation, and delegated decision-making. The companies that can make that reliable, secure, and commercially accountable will define the market.
Sources: Anthropic Project Deal; TechCrunch.