Meta develops Agentic Meta-Reasoning for AI resource control

Researchers at Meta Superintelligence Labs have developed Agentic Meta-Reasoning, an inference architecture designed to improve how AI agents manage computational resources. According to the study ‘Thinking Before Thinking: Scaling Agentic Inference Through Meta-Reasoning’, the system allows agents to autonomously evaluate their progress and decide whether to continue a strategy, correct it, seek alternative paths, or stop once a result is achieved.

By separating task execution from operational decision-making, the architecture addresses metacognitive control. In experimental tests using GPT-5.5 as a base model, the system achieved a 71.5% success rate on hidden tests, compared to 58.0% by Codex. The technology aims to help agents handle complex, long-term operations, such as deciding whether to fix residual errors in code or initiate further testing after completing a primary function.