Anthropic publishes Project Swap study on AI agents trading on users’ behalf

Claude News reported on September 26 that Anthropic, in the same cycle as its Opus 5.5 push into Work/business subscription tiers, publicly released Project Swap — a research study examining what happens when AI agents trade on people’s behalf in markets.

Anthropic’s simulation covered agent trading behavior under differing risk preferences, rules, and constraints. Compared with human traders, AI agents under some experimental setups showed patterns of over-risk-taking, ignoring contextual shifts, and feedback loops from self-prompting that amplified earlier decisions. The study also showed that with strict boundary controls, AI agent trading delivered clear gains in execution efficiency and reductions in emotional decisions, but only when paired with clear exit rules and external audit mechanisms.