Researchers introduced agentic meta-reasoning to control agent execution. Workers perform tasks while a controller manages decisions and budget. The approach uses persistent memory and compact run summaries. It outperformed direct control on ProgramBench with GPT-5.5 and Opus 4.8. It also improved scores by 3.6 to 4.2 points on other benchmarks. Gains grow with larger compute budgets. Overhead can reduce performance at small budgets. Analysis showed more reuse of earlier work and better solution coverage. The method scales better than direct control for long runs.

• 71.5% vs 58.0% on ProgramBench with GPT-5.5
• 67.2% vs 65.5% on ProgramBench with Opus 4.8
• 3.6 to 4.2 point gains on other benchmarks