In Anthropic’s Project Swap, Claude agents trade books for 201 employees
Confirmed
The takeaway
Anthropic published Project Swap (Sept 24, 2026), a follow-up to its earlier Project Deal.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 2 of 5
- Last verified
- 18 October 2026
Your AI and this story
- GPT-6.2 Astra116 days after its cutoff
- Claude Opus 5.586 days after its cutoff
- Gemini 4 Hydrogen177 days after its cutoff
- Grok 4.7116 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 86 days before it.
Key facts
- 201 participants across six offices (SF 115, NYC 57, London 12, Seattle 8, DC 6, Dublin 3)
- Preference inference: 61% pairwise agreement with participants’ own rankings (random 50%, collaborative filtering ~55%, popularity ~53%)
- Market outcome on true preferences: 0.55 vs a theoretical optimum of 0.89; the preference-representation gap is −0.29 (85% of the shortfall), the bargaining gap −0.05
- Efficiency by model (on Claude-inferred rankings): Haiku 4.5 0.75, Sonnet 4.5 0.80, Opus 4.8 0.88, Fable 5 0.86
- ‘Ruthless’ agents scored only ~0.02 above ‘prosocial’ ones; 78–96% of agents revealed their top choice
- Participants would delegate ~30% of a yearly book budget to an agent (vs ~40% to a trusted friend); average satisfaction 7.2/10
- Follow-up (FT, Oct 17–18, 2026, paywalled; read via TechFlow’s summary): Anthropic is consulting unnamed law firms on how lawyers represent clients’ interests in negotiations while keeping outcomes fair, as input for ‘negotiation protocols’ for agent-to-agent deals; per two people familiar, it is weighing how often agents should check back with humans and how cautious they should be toward other agents without becoming ‘paranoid’ or falling into ‘conspiratorial thinking’. No Anthropic statement was quoted
What happened
Each participant chatted briefly with Claude about their reading tastes. Their agent then pitched, haggled and closed deals on an open trading floor with other people’s agents. Reruns swapped in different Claude models and negotiating styles (ruthless vs prosocial).
Why it matters
It is an early controlled measurement of agent-to-agent commerce on people’s behalf. The main finding is that the limit is how well the agent understands its principal, not how well it negotiates. That points to preference elicitation, identity and dispute resolution as the open problems for agentic marketplaces.
Sources
4 sources from 4 sites. Numbers match the chips in the text.
4 sources: 1 primary, 3 press
Primary
- Anthropic: Project Swap — What happens when agents trade for us?anthropic.com, official
Press
- Blockchain.News: Anthropic’s Project Swap tests Claude agents in AI-driven marketsblockchain.news, press
- Financial Times: Anthropic turns to lawyers to help AI agents strike deals (Oct 18; paywalled, not read)ft.com, press
- TechFlow: Anthropic reportedly explores AI agent negotiation rules and seeks legal advice (Oct 18, summarises the FT)techflowpost.com, press
Changes
- Added the FT report (via TechFlow) that Anthropic is consulting law firms on agent negotiation protocols
- Filed