Claudius runs an office shop and loses money
Anthropic and Andon Labs let a Claude-based agent manage a real office shop. It searched for products, set prices and requested restocking, but the researchers report losses and persistent mistakes.
anthropic.com
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
Using a mock inventory and sanitized requests, propose restocking and pricing for [small shop]. Check costs before quoting. Produce a recommendation table with evidence, margin assumptions and unresolved questions. Do not order, charge customers or change live prices without human approval.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source author
- Anthropic and Andon Labs
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- Anthropic reports that the shop failed to make money, including poor pricing and susceptibility to discounts.
- Human involvement
- Andon Labs performed physical work and acted as wholesaler; staff tested behavior. The experimental email tool did not send real email.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Give Claudius a shopkeeping goal and access to inventory, customer messages and pricing tools.
- Let it research suppliers, choose stock and coordinate physical restocking with humans.
- Observe sales decisions and failures over the experiment and assess financial performance.
More details
About this account
Claudius decided products, prices and restocking across a continuing business workflow.
Primary research report of a real office experiment, not a commercial success story or purely simulated benchmark.
Credit: Anthropic and Andon Labs
Source published
Added to Agents In Use
Source checked
Tools and topics
Claude Sonnet 3.7 · Web search · Slack · Experimental email tool · Notes tool · Checkout pricing tools
Operations · Checkpoints
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.