AI & Tech··1 sources verified
Study Finds AI Agents Lose Learned Rules When Handing Off Written Summaries
Summarised by Relevant News AI · Read time: 3 min
Researchers tested six models across 56 interactive worlds with hidden rules and found that a chain of agents passing 300-character notes learns far less than a single agent that keeps its full history. After eight handoffs the chain performed about where the single agent stood after four lives. Compressing notes and repeated compression caused most of the loss, and writing each note from the full record of earlier lives recovered 70% to 82% of it.
Why it matters: Teams building multi-agent systems or agents that span context windows should know that summary-based handoffs silently discard what the agent has learned, so handoff design needs to draw on the full history rather than the prior note.
All sources