Announcement
A proposed framework lets AI agents plan their own computation
The paper’s authors describe a controller that weighs next steps against remaining compute and directs workers using persistent memory.
TLDR
The researchers propose a controller that tracks established work, weighs possible next steps against its compute budget and dispatches workers. They report 71.5% on ProgramBench with GPT-5.5, versus 58.0% for Codex. On other benchmarks, they report gains of 3.6 to 4.2 points over direct control, averaged across three frontier models. They note that the framework’s overhead can hurt at small budgets.
Combined views
—
1 Source, first seen ago
