Agents that negotiate are a legal product before they are a technical one, and Anthropic appears to be treating them that way. The Financial Times reports, citing sources, that the company is consulting law firms on deal negotiation tactics as it develops agents that bargain on a person’s behalf and interact with other people’s agents. Source: https://www.techmeme.com/261018/p2#a261018p2
Key facts
| Fact | Detail |
|---|---|
| Question answered | Anthropic Is Teaching Agents to Negotiate - With Lawyers in the Room |
| Short answer | The Financial Times reports that Anthropic is consulting law firms on deal negotiation tactics as it builds agents that negotiate on a person’s behalf and deal with other people’s agents. |
| Reading time | ~4 min |
| Last updated | 2026-10-19 |
The reported framing is the useful part: tech companies are being forced to confront the rules of negotiation when machines act for humans. Not the rules of haggling - the rules of authority, disclosure and liability.
What is actually reported
- Who: Anthropic, per Cristina Criddle’s FT report, citing sources.
- What: consulting law firms on negotiation tactics, in the context of agents that negotiate for users and against other agents.
- Why it is framed as a rules problem: the FT’s own summary puts it as agents needing to learn “how to interact in negotiations, representing their clients while honouring fair outcomes.”
Coverage of the piece has added detail - a reported internal exercise where staff agents transacted with each other, and consultation with a large law firm. Those specifics come from summaries of the FT report rather than the FT text itself, so treat the numbers as reported rather than audited.
Negotiation is mostly not about the number
Anyone who has negotiated seriously knows the price is the last thing discussed. The substance is:
- Authority. Can this person agree, or must they check? An agent without an explicit mandate is either useless or dangerous, depending on which way it errs.
- Disclosure. What must you reveal, and what may you conceal? The line differs by jurisdiction and by contract type, and it is not something a model can infer from a corpus.
- Fair dealing. Some tactics that work are unlawful or voidable - misrepresentation, duress, unconscionability. An agent optimised for “win the deal” will find them.
- Formation. When exactly is a deal closed? A click, an email, a JSON payload? If the answer is ambiguous, so is enforceability.
Law firms are the right consultants for this because the constraints are not in the training data as rules; they are in case law, professional conduct rules and the specific contract at hand.
The agent-to-agent problem is different
Two humans negotiating have separate principals, separate interests and a shared understanding that they are adversarial. Two agents negotiating introduce failure modes that neither party controls:
- Tacit collusion. If both agents are trained toward the same “fair outcome” objective, they may converge on a price that is fair to each other and bad for both principals. Nobody agreed to collude; the objective did it.
- Common vendor. If both sides rent their agent from one of three providers, the only party present on both sides of the table is the vendor. That observation circulated alongside the FT report and it is the sharpest thing said about this whole category.
- Accountability gap. If your agent conceded a term you never saw, and their agent accepted it, which human is bound? “The agent did it” is not a legal defence; it is a description of your exposure.
None of these are solved by better prompting. They are solved by mandates, audit trails and limits - which is presumably why lawyers are in the room before the product is.
If you let an agent negotiate for you
You do not need to wait for Anthropic’s product to run into these questions. Any agent that sends messages on your behalf is already negotiating.
- Write the mandate down. A one-page brief: what you want, what you will accept, what is off-limits, who must approve. Give it to the agent as a document, not a vibe.
- Set walk-away limits in code, not prose. A hard floor that triggers human review is more reliable than an instruction to “be reasonable.”
- Require approval above a threshold. Below it, the agent can move; above it, a human signs. The threshold is the whole design.
- Log every concession. Not the final deal - the sequence. What you gave up, in what order, in exchange for what.
- Decide your disclosure posture. Whether your counterparty is told they are talking to an agent is a choice with legal consequences in some jurisdictions. Make it consciously.
Where a desktop agent fits
The negotiation case is the sharpest argument for local, inspectable agents. If an agent is going to commit you to something, you want to see the draft before it sends, know which files it read, and be able to stop it mid-thread. A cloud service that “handles it for you” is convenient right up until it agrees to something.
MOOGH runs on your machine, shows the actions it takes and keeps approval with you - which is the minimum for anything that speaks on your behalf: [download MOOGH].
FAQ
Is Anthropic shipping a negotiation agent? The FT reports development work and consultation with law firms. No shipped product with published terms is described in the material summarised here.
Why not just let the model read contract law? Because the operative rules are jurisdiction-specific, updated by case law, and often about professional conduct rather than statute. Retrieval helps; it does not substitute for a mandate.
What is the single most important control? The approval threshold. An agent that can commit you to unlimited terms is not a tool, it is a counterparty.
Does this affect individual users or only enterprises? Both. Anyone who lets an agent reply to a quote, accept terms, or book a service at a negotiated rate is in scope - the amounts are smaller, the liability logic is the same.
Related reading: