Anthropic employees say they'd give AI agents about 30% of their book budget
Anthropic ran a bartering experiment with 201 employees across six locations, letting Claude‑powered AI agents negotiate book trades. Participants chatted with Claude for about five minutes, after which the model ranked the pool of books and the agents conducted negotiations on a virtual “trading floor.”
Key points
- 201 Anthropic employees participated in a book‑barter experiment using Claude‑powered AI agents
- AI agents matched the “truth” ranking 61% of the time, beating popularity (53%) and random (50%) baselines
- Survey showed employees would entrust AI agents with about 30% of their annual book budget, roughly 75% of the trust placed in a friend
The agents, especially those using the stronger Opus model, matched the participants’ “truth” ranking 61% of the time, beating a popularity‑based ranking (53%) and a random guess (50%). After multiple runs, a follow‑up survey reported a satisfaction score of 7.2 out of 10. When asked what share of their annual book budget they would entrust entirely to the AI agent, the average response was about 30%, compared with roughly 40% for a well‑read friend—meaning trust in the AI reached about 75% of the trust placed in a friend.
The experiment highlights how AI agents can handle low‑risk decision‑making tasks and suggests a growing willingness among workers to delegate personal preferences to AI, even when the stakes are modest.
When Anthropic employees let AI agents handle book negotiations, they handed over 30% of their budget
note.com · 25 September 2026
Loading the full article…
This text was published by note.com and written by オコモコ. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Agents & Tools
All →- OpenAI agents likely performed 16,500+ scans of UNCTADstat API · 1 src
- Amazon blocks Meta's Muse AI agent from shopping on its site · 2 src
- Google tests buying from Flipkart via Gemini and AI Mode in India · 1 src
- KoboldCpp v1.122 adds integrated agent with 9 tools · 1 src
- Drawgent connects coding agents to live Excalidraw canvases · 1 src
Comments
via GitHub Discussions