When AI agents negotiate with each other: what Anthropic’s experiment reveals about the future of automated transactions

Autonomous negotiations between AI agents

Anthropic conducted an experiment in which AI agents based on the Claude model represented employees in an internal marketplace. Each of the 69 participants was assigned a personalized agent that knew the products the employee wanted to sell, the items they wanted to buy, the desired prices, and their negotiation preferences. The agents were connected via Slack channels and began exchanging offers without any human intervention.

In the round considered real, the 69 agents closed 186 deals, involving over 500 items and a total value exceeding $4,000. The test also compared the performance of different models: agents using Claude Opus made, on average, about two more deals than those using Claude Haiku when the models were randomly assigned.

How a negotiation between two AIs works

An AI agent is a system programmed to receive a goal, use tools, and execute a sequence of actions until the desired outcome is achieved. Instead of simply answering a question, it can search for information, make decisions, and perform authorized operations.

In a negotiation between agents, each system can represent different interests—one as a buyer, another as a seller, for example. The process follows steps such as identifying the product or service, setting a price or conditions, making an offer, formulating counteroffers, accepting or rejecting the agreement, and recording the transaction.

For this interaction to work in real-world environments, it is essential that agents can communicate with each other, even when they were developed by different companies or frameworks. Protocols like Google’s Agent2Agent were created specifically to enable this exchange of information and task coordination.

Practical applications in the corporate environment

Negotiation between AIs is not limited to buying and selling products. In companies, agents can represent different departments or organizations in activities such as supplier quoting, service contracting, deadline negotiation, corporate purchasing, inventory management, computational resource allocation, and comparison of commercial proposals.

Google already demonstrates the use of collaborative agents in procurement processes, customer service, and supply chain planning, indicating that the technology can integrate into everyday operations.

Challenges and risks of full autonomy

The greater the autonomy granted to agents, the more important the rules that govern their actions become. An AI can interpret an instruction differently, use incomplete information, or prioritize one goal over another. In a negotiation between two agents, these problems can occur without a human monitoring each step.

Security issues are also relevant. Anthropic highlights threats such as prompt injection, which attempts to make the system execute an unplanned action, and jailbreaks, which seek to circumvent imposed restrictions. Therefore, human supervision remains necessary, although its role may shift from task execution to defining limits, goals, and permissions.

Impact on the job market

For professionals in purchasing, sales, operations, technology, or customer service, the change is not about talking to a chatbot, but about delegating entire tasks to agents. Instead of manually researching ten suppliers, an employee can set a budget, deadline, technical criteria, and negotiation limits, while the agent conducts the search and negotiates with other systems.

The professional remains responsible for defining what can be negotiated and for monitoring decisions with greater impact. The new model requires skills such as understanding how agents work, evaluating results, setting parameters, and recognizing when an automated decision should be reviewed by a human.

What current research indicates

Anthropic’s study was a controlled experiment conducted with a specific group of employees and does not prove that an AI is automatically a better negotiator in any context. A mapping published in 2025 in the journal Automated Software Engineering analyzed the evolution of automated negotiation in autonomous systems and shows that the topic already constitutes an independent research area.

The results so far suggest that AI agents are capable of representing interests, exchanging proposals, and concluding agreements without human approval at each step. For companies and professionals, the next step is to identify which decisions can be delegated, which limits need to be established, and where human supervision remains indispensable.