Anthropic is exploring AI agents that could haggle, settle disputes and bargain for people, while seeking legal advice on how such negotiations should work. The reported details point to a practical challenge for autonomous AI: setting limits on what an agent may agree to, and when it must check back with its user.
Anthropic Watch analysis
What happened
The Financial Times reports that Anthropic has approached law firms for advice on negotiation tactics and is considering how often an agent should consult its user during a negotiation. One idea is to give agents signed documents setting out a user’s instructions and rules.
The report also describes a September experiment in which Anthropic employees had agents trade books on a virtual trading floor. Agents told to act ruthlessly finished slightly ahead; cooperative agents made sacrifices so every participant ended up with a book. Anthropic declined to comment to the FT, according to The Daily Star’s account.
Why it matters
A negotiating agent is not just drafting a message. It may need to make trade-offs, judge another agent’s intentions and decide whether a deal falls within its brief. Those decisions raise questions about authority and fair dealing before anyone gets to the fine print.
Our read
The interesting shift is from agents that recommend a response to agents that may commit someone to a bargain. Signed instructions sound like a sensible starting point, but the hard part will be making those limits clear and enforceable when negotiations get messy.
What to watch
- Whether Anthropic publishes a product, prototype or further results from its experiments.
- How agents would verify each other’s authority and respond to conflicting instructions.
- Whether the legal advice leads to concrete safeguards or stays at the discussion stage.
Discussion spark: Should an AI agent be allowed to accept a deal within written limits, or should a person have to approve every agreement?
Sources and evidence
- Anthropic consults lawyers on how AI agents should negotiate: report – The Daily Star (23 October 2026, 07:46 UTC)
Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.