AI agents can bargain. Can they represent you?

What happened

On September 24, Anthropic published Project Swap, a controlled market in which employee participants briefed Claude agents to exchange books. Among the 188 participants who supplied rankings, the agents’ inferred preferences agreed with people on 61% of book pairs. Researchers attributed most of the gap from the best possible allocation to imperfect preference information, rather than bargaining itself.

Sources: Anthropic

What the evidence supports

This was a small internal experiment, not a recruitment or procurement trial. Employee volunteers are not representative of the public; the study did not measure job losses. Its result is evidence about this trading setup, not a general success rate for autonomous negotiation.

Sources: Anthropic

Our interpretation

For a business testing delegated decisions, we would separate three questions: Did the system obey the brief? Was the brief complete? Was the final outcome useful? Combining those into one success score can conceal a basic problem: a technically compliant decision may still disappoint the person responsible for it.

Consider a hypothetical sales assistant authorized to negotiate a subscription. A lower price could look like a win while the buyer actually values cancellation flexibility or implementation support more. In recruitment, a higher salary may be less attractive than a predictable schedule. These are examples of evaluation criteria a team could specify, not outcomes tested in Project Swap.

The occupational implication is a question to investigate: will employers value people who translate ambiguous needs into reliable decision processes? A convincing answer requires workplace evidence. For now, we would treat the ability to define acceptance criteria, inspect trade-offs, and explain exceptions as a useful portfolio project for sales operations or HR staff experimenting with AI.

What to do next

Choose one recurring, low-stakes decision from your work. Write down an acceptable result, two unacceptable compromises, and the information needed to tell them apart. Run a few historical examples without executing any agreements. Compare the assistant’s choices with the decision you would defend to a colleague.

Keep a short error log: missing context, wrong priority, factual mistake, or unclear instruction. Use it to decide what further testing would be worthwhile. A useful demonstration should show where intervention was needed as well as where the assistant helped. That gives a prospective employer or manager something concrete to evaluate beyond a fluent transcript.