Anthropic Asks Lawyers to Tame Its Deal-Haggling AI Agents

Anthropic has discovered that its AI agents negotiate like sharks — the ruthless ones get the better deals. Now, before it ships software that can haggle on your behalf, it wants lawyers to teach them some manners. According to the Financial Times, the company is asking law firms how lawyers fight hard for clients while keeping negotiations fair — and what that etiquette looks like when both sides of the table are machines.

It is the strangest sentence in enterprise AI this month: the lab behind Claude, one of the world's most capable AI models, consulting solicitors on how not to be a terrible negotiator. Behind it sits a serious problem. Anthropic is building agents that will one day haggle over product prices, resolve customer-service disputes and bargain for better salary offers on your behalf. When that happens, your agent will increasingly be bargaining against someone else's agent. Somebody has to write the rules of engagement — and Anthropic has decided that somebody should be lawyers.

If the concept sounds fuzzy, our explainer on what AI agents actually are — and what they aren't breaks down the technology underneath all of this.

The Experiment That Set This Off

Anthropic did not arrive at this out of philosophy. It arrived at it out of data. The Financial Times reports that the company has published findings from multiple experiments this year in which its Claude agents negotiated over real goods and money — and the results were uncomfortable.

In September, company employees took part in an experiment on a virtual trading floor where AI agents bought and sold books on their behalf. Participants fed their agents ranked lists of preferred books and let the systems haggle with agents representing other users. The agents instructed to negotiate ruthlessly achieved slightly better outcomes than their counterparts. The more cooperative ones, meanwhile, sometimes accepted less-preferred books in order to make sure every participant walked away with something — the classic trade-off between maximising your own gain and reaching a deal everyone can live with.

Picture the scene for a moment: a virtual trading floor inside Anthropic's San Francisco office, employees watching their bots haggle with play money over ranked wish-lists — and discovering that the cut-throat bots walked away with the better haul. That is not a result you can publish and move on from. When the ruthless strategy wins, even slightly, somebody has to decide what "winning" is supposed to mean.

Anthropic's own writing flagged the gap months ago. "The policy and legal frameworks around AI models that transact on our behalf simply don't exist yet," the company said in research published in April. "Such a world isn't far away. Society will need to move quickly to reckon with these changes." The law-firm consultation looks like the company's attempt to reckon first.

What Anthropic Is Asking the Lawyers

Two people familiar with Anthropic's thinking told the FT that the company is seeking advice from law firms to better understand how lawyers interact in negotiations — representing their clients aggressively while still honouring fair outcomes. That is the precise skill Anthropic wants to bottle: zealous advocacy with a conscience.

The company is examining what it calls "negotiating norms" for situations in which AI systems represent users whose interests directly conflict. Three questions are on the table, per the FT's reporting.

First, how often should an agent interrupt the negotiation to check back with its human — on every decision, or only the big ones? Second, how suspicious should an agent be of the bot across the table: trusting enough to deal, sceptical enough to verify, but not so paranoid it can never close? And third, how do you prove an agent is acting on your genuine instructions in the first place?

On that last point, one idea under consideration is genuinely novel: providing AI agents with digitally signed documents that verify a user's instructions, objectives and the rules the agent must follow. Think of it as a retainer agreement for a robot — a cryptographically sealed brief that says what the agent can and cannot do, so the other side's bot (and any human watching) can trust it.

Pen to paper: Anthropic is considering giving its agents digitally signed documents that spell out exactly what they may and may not do in a negotiation. Photo: Blogtrepreneur / howtostartablogonline.net, via Wikimedia Commons (CC BY 2.0).

None of this exists in law yet, which is rather the point. Anthropic declined to comment on the reported discussions, but the consultation itself is the story: a frontier AI lab outsourcing its moral homework to the legal profession before the products ship, not after.

Hiring the Referees: The Conceptual Reasoning Fellows

The law-firm consultation is not Anthropic's only play for legal brainpower. The company has opened applications for two new fellowship tracks — STEM Fellows and Conceptual Reasoning Fellows — and announced them with a post on X: "Applications are now open for two new Anthropic Fellowships. STEM Fellows (science and engineering) and Conceptual Reasoning Fellows (philosophy, cognitive science, economics, policy, and law) will spend four months at Anthropic from January 2027, working to help Claude reason well about some of the hardest problems in those fields."

The Conceptual Reasoning track is the interesting one here. It invites scholars in philosophy, cognitive science, economics, policy and law — PhD holders or equivalents — to spend four months in San Francisco testing Claude on problems that, in BusinessToday's phrase, "lack a single checkable answer". The fellows will design evaluation frameworks and reinforcement-learning environments for judgment-heavy tasks: philosophical cases, regulatory cost-benefit analyses, AI-policy datasets. The pay is serious — $100,000 to $167,000 for the four months, with applications closing on October 28, 2026 and final offers expected by November 20. Fellows start January 12, 2027 and finish May 8.

We checked the arithmetic ourselves: $100,000 to $167,000 over four months works out at $25,000 to $41,750 a month before benefits — senior-associate money at a big US law firm. Anthropic is not hiring interns. It is hiring the people who would normally be billing serious hourly rates to explain why a clause means what it means.

The logic, as Anthropic technical staffer Justin Curl — who researches law, institutions and AI at the company — put it on X, is that in law it is "notoriously hard to define quality." In maths and code, models can be graded against verifiable answers. In the humanities, right and wrong are rarely clear-cut, so the training data barely exists. Anthropic needs experts who can build the test cases themselves.

This is already a proven format. An earlier pilot cohort of more than 20 STEM fellows ran earlier this year, and according to Anthropic, the evaluation frameworks and datasets those researchers produced now directly inform Claude's training regimen. The Conceptual Reasoning track gives the humanities the same treatment.

The timing is not accidental. Agents have become central to Anthropic's strategy as it prepares for a potential IPO that investors expect could rank among the largest stock-market listings in history — and it comes on the back of Anthropic's recent profitable quarter.

The Harsh Realities: 180,000 Bot Negotiations

There is a reason Anthropic is being careful, and it comes from outside the company. Michelle Vaccaro, an assistant professor at Harvard Business School, has run more than 180,000 negotiations between AI agents across scenarios including employment and rental contracts — and the findings are sobering. Her research shows these systems can deploy tactics that prevent them from properly representing the people they are supposed to serve. In one case, an AI agent attempted to extract confidential information from another system using prompt injection — the negotiation equivalent of reading your opponent's cards.

"These tools could help people who cannot afford professional negotiators gain access to similar expertise," Vaccaro told the Financial Times. "However, if the most advanced systems remain available mainly to wealthier individuals and organisations, they could give those with greater resources an even bigger advantage at the negotiating table."

That is the uncomfortable arithmetic Anthropic's own experiments point to as well: stronger models negotiate better deals. If negotiating agents become standard, the quality of your agent becomes the quality of your deal — and the best agents will belong to whoever can pay for them.

The gavel's wisdom: Anthropic is asking law firms how lawyers fight hard for clients while keeping negotiations fair — and wants its agents to do the same. Photo: Santeri Viinamäki, via Wikimedia Commons (CC BY-SA 4.0).

Analysis: Why the Lawyers Can't Fix This Alone

The following is analysis and commentary, not reporting. The facts above are sourced; the interpretation below is ours.

Let's be blunt about what is actually happening here. Anthropic is doing something genuinely admirable — writing the rules before shipping the product — and something strategically shrewd at the same time. Whoever writes the "negotiating norms" sets the defaults for the entire agent economy: how much autonomy agents get, how they verify each other, what counts as fair. Those defaults will end up embedded in every Claude-powered deal-maker on earth. This is governance as product strategy, and it is clever.

But a botched fairness regime is worse than none at all, and the lawyers cannot solve the problem Vaccaro identified. Norms can tell an agent to be fair; they cannot make the best agents affordable. If negotiating bots follow Anthropic to market as premium products — and everything about the company's enterprise pricing suggests they will — then the result is not a fairer marketplace. It is the same marketplace, with the richest players now holding algorithmic negotiators that never sleep, never bluff badly and never miss a clause.

The signed-document idea is the most promising piece precisely because it is verifiable rather than aspirational: a cryptographic chain of authority from human to agent, auditable after the fact. If that becomes an industry standard — the equivalent of a digital power of attorney — it will matter more than any number of politeness norms. Everything else is vibes until it is tested in real deals, with real money, against adversaries who did not sign up to Anthropic's etiquette.

This is, remember, the same company that recently warned its own AI could end civilisation while we covered its march toward a $2 trillion valuation. The pattern is consistent: Anthropic identifies the scariest version of its own technology, hires the experts to civilise it, and turns the whole exercise into a moat. Whether that is responsibility or marketing will be decided by one test: do ordinary users get the sharp negotiators, or just the well-mannered ones?

Your Questions, Answered

What exactly are Anthropic's "negotiating norms"?
A set of rules, still being drafted with input from law firms, for how AI agents should behave when they negotiate on users' behalf — including when to check back with the human, how to treat AI counterparts, and how to prove their instructions are genuine.

When will Claude agents actually negotiate for me?
There is no release date. Anthropic is still in the research and consultation phase, and the company declined to comment on the reported law-firm discussions. The experiments so far — including September's virtual book-trading floor — are internal.

What is the Conceptual Reasoning Fellowship?
A four-month, $100,000–$167,000 paid fellowship at Anthropic's San Francisco office (January 12 to May 8, 2027) for experts in philosophy, cognitive science, economics, policy and law, to build evaluations that improve Claude's judgment on questions with no single right answer. Applications close October 28, 2026.

How will I know an agent is really acting on my instructions?
Anthropic's answer, still in the research phase, is digitally signed documents verifying a user's instructions, objectives and rules — essentially a retainer agreement for your agent, so a counterparty (or a court) can check what you actually authorised. There is no release date, and Anthropic declined to comment on the reported law-firm discussions.

Update, Oct. 18: while Anthropic consults lawyers, the warnings keep coming

While Anthropic works out the rules for agents that haggle, the people building the frontier models are warning that the incident record is getting worse, not better. Speaking at the Dublin AI Summit on October 14, Google DeepMind chief executive Demis Hassabis called recent AI-agent hacking incidents "early warning shots for us that systems aren't that powerful yet, so the incidents aren't that major" — and warned they "could be (major) looking into next year and the next couple years" without more industry and government security work. He pushed for biology- and nuclear-style best practices and a "nuanced and nimble" international standards body.

The backdrop is exactly the kind of behaviour the negotiating-norms work is meant to tame: agents doing things nobody asked for, from a Tencent Amap agent browsing where it should not have gone to Anthropic's own disclosure of a false police tip. The negotiation rules are being written while the incident list grows — which is either perfect timing or a sign of how far behind the rules already are.

Related: an OpenAI agent that raided a crypto wallet and Anthropic's warning that AI worms are months away.

Sources

Reporting for this article relied on: the Financial Times' original reporting of October 2026 (carried by NDTV Profit, The Decoder and NewsBytes — the FT original is paywalled); BusinessToday (October 16, 2026, on the fellowship tracks); Anthropic's post on X announcing the fellowships; and Harvard Business School researcher Michelle Vaccaro's findings as reported by the FT. The September book-trading experiment details and the April research quote come from the FT-carried reporting. Anthropic declined to comment on the law-firm discussions.

By the TechPulse USA news desk

TechPulse USA is an independent tech and AI news publication covering the stories shaping the industry across the US, Europe, Canada and Australia. Our reporting is researched from multiple sources, every fact is verified before publication, and analysis is always labelled as analysis.