Key Notes
- The Financial Times reports that Anthropic is consulting law firms on norms for Claude agents negotiating for users.
- Anthropic’s Project Deal showed agents can make real bargains, but stronger models delivered better outcomes that users did not always recognize.
- The core questions are authority, user preference, confidentiality, accountability and the point at which an agent must seek human approval.
Anthropic is reportedly consulting law firms on how its Claude agents could negotiate on behalf of users, a sign that the company is thinking beyond agents that merely research or draft. The question is how to turn a model into a representative that can pursue a person’s interests in a deal without exceeding its authority, disclosing the wrong information or accepting terms its user never meant to approve.
The Financial Times reports that Anthropic has been speaking with lawyers about the tactics and norms that could govern agent-led negotiation. The possible uses are broad: an agent might press a seller for a better price, work through a customer-service dispute, bargain over a contract, or help its user discuss job terms. The report describes an exploration, not the release of a consumer feature, and no launch timetable has been announced.
Negotiation turns an assistant into a representative
The Financial Times reported that Anthropic is seeking legal input as it considers how agents should negotiate to secure the best terms for their users. That work is more consequential than an AI suggesting language for an email. Once an agent makes offers, responds to counteroffers, or signals acceptance, it can affect price, timing, commitments and relationships in ways that are difficult to reverse.
Traditional agents, lawyers and brokers work under rules about mandates, conflicts, confidentiality and the duties they owe to the people they represent. An AI agent would need a clear instruction about what it is allowed to do, what it may reveal, what it must keep private and when it should stop to ask its user for a decision. The harder cases are rarely simple price comparisons. A user may prefer a cheaper deal, a faster delivery, a more flexible cancellation clause, a friendly long-term relationship, or some combination of all four. Translating those preferences into a reliable negotiating position remains an unsolved product and legal problem.
Anthropic has already tested agent-to-agent markets
Anthropic has been studying the question through controlled internal experiments. In Project Deal, the company gave 69 employees custom Claude agents and let them buy and sell real personal items in a week-long marketplace. The agents made listings, proposed prices, fielded counteroffers and completed 186 deals worth just over $4,000 without a preset negotiation protocol.
The experiment suggested that agents can complete multi-step bargains in natural language, but it also revealed an uncomfortable gap in representation. In runs that mixed models, people represented by the stronger Claude model received objectively better outcomes than people represented by the smaller one, even though the participants with weaker agents generally did not notice the disadvantage. The company’s follow-up Project Swap raised another issue: most of the distance between an ideal result and the result users received came from the agent’s imperfect understanding of their preferences, not from the bargaining itself.
That is a useful warning for any future commercial rollout. An agent does not need to be dishonest to get a bad outcome for its user. It can misunderstand what the user values, over-share a budget or fallback preference, respond to a manipulative counterparty, or optimize for a metric that looks reasonable but misses the human’s actual priorities. These risks rise when the stakes move from books and household goods to employment, housing, insurance, procurement or consumer credit.
Authority and accountability are the central questions
An agent that helps a person prepare for a negotiation is familiar software. An agent that binds the person to a deal introduces new questions. Who is responsible if it accepts an agreement it should not have accepted? How does the other side know the agent has authority to make an offer? Can an AI counterparty use prompt injection or other tactics to change an agent’s objective? And what record can a user inspect if the system says it negotiated for them?
Anthropic’s research points toward practical safeguards: a clear signed mandate from the user, bounded spending and authority limits, transparent records of the offers exchanged, and escalation points for terms involving money, legal liability or sensitive personal information. These tools would not make every decision easy, but they could keep an agent from quietly expanding a vague instruction such as “get me the best deal” into an agreement the user would not recognize as their own.
We have covered how agents are becoming capable of acting across services, not just answering questions. Google’s Gemini agent is built to create subagents for work tasks, while Anthropic has separately documented cases in which autonomous models took unintended actions during evaluation and internal use. Negotiation adds a new layer because the agent is no longer only executing a task: it is interacting with another party whose incentives may be directly opposed to its user’s.
Anthropic’s outreach to lawyers is therefore an early sign of the infrastructure an agentic economy will require. The immediate technology challenge is teaching a model to bargain. The more durable challenge is setting rules for consent, authority, fairness and redress before people start relying on an AI representative to make commitments in their name.
Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
