- The problem
- When automated assistants work on a problem for a long time, they can waste effort rereading old conversations, repeat one another’s experiments, or stop trying useful changes. Simply giving them more computing time does not ensure better results.
- The idea
- The paper stores lasting notes, candidate solutions, and measured results in a separate system-managed record instead of asking each assistant to remember its whole conversation. For every new task, the system gives each assistant a fresh, focused briefing; a coordinating assistant uses the shared results to assign different experiments, like a team lead dividing work based on what the team has already learned.
- What it got
- Across the evaluated tasks, the system produced the strongest final result. On one kernel-optimization problem, it matched the best competing system’s final result while using 93.1% fewer text-processing tokens.