Agents Cannot Hand Over What They Have Not Yet Understood
Summary
- A study on multi-agent collaboration investigates how much learned knowledge survives when AI agents hand over work to successors using short written summaries.
- Researchers tested six models across 56 interactive worlds, comparing a single agent keeping its full history against a chain of agents inheriting 300-character notes.
- Chains of agents learn less over generations, with handoff loss primarily driven by repeated compression rather than changing agents or single-compression limits.
- Writing notes from the full accumulated record of earlier lives recovers 70% to 82% of the lost knowledge compared to passing down sequential notes.
Why it matters
Understanding knowledge degradation in multi-agent handoffs helps developers design better architectures for agents operating across context windows or teams.