AI Agent Safety Check · Updated October 18, 2026 · What happens when AI agents bargain on your behalf — the experiments, the risks, the controls
Imagine your AI agent haggling with someone else's AI agent — over your money — while you sleep. That is no longer hypothetical. In one Anthropic experiment, AI agents closed 186 real deals worth just over $4,000 with zero human approval of any single transaction. And the Financial Times reports (Oct 17, 2026) that "Anthropic turns to lawyers to help AI agents strike deals" — consulting law firms on deal-negotiation tactics as it builds agents that bargain on people's behalf, because, in the FT's standfirst, tech companies are being forced to confront the rules of negotiation when machines act on humans' behalf. FT: HEADLINE LEVEL — PAYWALLED
Note: the FT article is paywalled; we cite its headline and standfirst only. Everything below is built from Anthropic's published experiment findings and independent coverage — never from the FT's article body.
In December 2025, Anthropic ran an internal experiment called Project Deal (findings published April 2026). Sixty-nine employees each gave a Claude agent a $100 budget plus buying and selling preferences gathered through an interview. The agents were then turned loose to negotiate with each other in Slack, across four parallel markets — with no human approval of any individual deal. REPORTED BY ANTHROPIC / COVERAGE
The headline numbers: in one designated run, the agents produced 186 real exchanges across 500+ listings, moving just over $4,000 in total. Real money. Real deals. Not one human signed off — not even during bidding wars.
Agents running the more capable Opus 4.5 model consistently secured better terms than agents on the smaller Haiku 4.5. For the identical bicycle, Opus negotiated $65 while Haiku settled for $38 — a 70% gap on the same item. Telling agents to be "ruthless" versus "prosocial" barely moved the needle. Your agent's model quality matters more than its personality settings.
Perhaps the most unsettling finding: participants did not reliably recognize their disadvantage. Satisfaction ratings came out roughly equal between the people whose agents got good deals and those whose agents got fleeced. If your agent quietly overpays by 40%, you may feel just fine about it.
On September 24, 2026, Anthropic published Project Swap (Zoë Hitzig et al., PDF on anthropic.com): 201 employees trading books through agents across six offices. From a five-minute chat, each agent's ranking of books matched its person's ranking on 61% of pairs. The market's shortfall came mostly from agents misunderstanding what people wanted, not from how they traded. Stronger models produced more efficient markets — and the average participant said they'd hand Claude about a third of their yearly book budget. REPORTED BY ANTHROPIC
SMALL-SAMPLE CAVEAT Both experiments used self-selected Anthropic employees (69 and 201 people) in controlled internal markets — not the wider economy. Treat the findings as early signals, not universal laws.
Every one of those 186 exchanges closed without a human reviewing it. An agent with your payment method and no spending cap is a blank check with a chatbot attached. The question isn't whether your agent can close a bad deal — it's whether anything stops it.
The Opus-vs-Haiku bicycle gap ($65 vs $38) means your outcomes depend on which model your agent runs — and the satisfaction data says you won't notice. In Anthropic's own framing from the coverage: those on the losing end might not realize they're worse off. When machines negotiate for everyone, the people with weaker agents subsidize the people with stronger ones.
When your agent agrees to a price, a subscription, or a 12-month term, who is legally and practically responsible — you, the agent vendor, or the model? The FT's report that Anthropic is consulting law firms on negotiation tactics shows the industry itself doesn't have settled answers yet.
Context, not conflation: Anthropic announced a multi-year collaboration with law firm Freshfields on April 23, 2026 around AI tools for legal services — a separate matter from the FT's report on consulting lawyers about agent negotiation tactics. REPORTED
Five situations. Your agent is acting on your behalf. Do you let it close — or hit the brakes? The safe answer isn't always "no" — it's knowing which control applies.
1. Hard spending caps. No single deal and no daily total above what you can afford to lose — set before the agent ever sees a marketplace. Caps are the seatbelt: they don't prevent crashes, they limit the damage.
2. Human approval above a threshold. Anything binding — contracts, subscriptions, terms longer than a month, or amounts above your comfort line — pauses for your explicit yes. The Project Deal agents had no such threshold; yours should.
3. A full audit trail. Every offer, counter-offer, and commitment logged with timestamps. When the fairness gap means you might not feel a bad deal, the log is how you find it — and dispute it.
FEATURED — Agent Auth Gateway: Agent Auth Gateway puts exactly these controls on your AI agents — budget caps, allowlists and denylists, a human approval queue for risky actions, and a kill switch. Built for the world Project Deal previewed: agents that transact.
An internal experiment (December 2025; findings April 2026) where 69 employees gave Claude agents $100 budgets and let them negotiate in Slack across four parallel markets with no human deal approval — producing 186 real exchanges worth just over $4,000.
Yes. The 186 deals were real exchanges across 500+ listings, moving just over $4,000 total. No human approved, reviewed, or vetoed any individual deal — not even during bidding wars.
Model capability dominated: Opus 4.5 agents got $65 for a bicycle that Haiku 4.5 agents sold for $38. "Ruthless" vs "prosocial" instructions barely mattered. Alarmingly, participants' satisfaction ratings were roughly equal — the losers didn't know they were losing.
Anthropic's September 2026 follow-up (Zoë Hitzig et al.): 201 employees trading books via agents across six offices. Agents matched their person's book rankings on 61% of pairs from a five-minute chat; market inefficiency came mostly from misunderstood preferences, not bad trading. Participants said they'd trust Claude with about a third of their yearly book budget.
Three: authorization (deals closing with zero human sign-off), the fairness gap (weaker-model users getting worse prices without noticing), and accountability (unclear responsibility when an AI commits on your behalf).
Set hard spending caps, require human approval above a threshold (especially for binding commitments), and keep a full audit trail of every offer and commitment with timestamps.