Meta researchers describe a separate controller for long-running AI agents: it assesses progress and chooses what work to do next. In the paper’s main settings, the system scored higher than matched direct control in all 12 comparisons across four benchmarks and three models. The preprint evaluates a research system, and its advantage varied with the model-call budget.
The approach, called agentic meta-reasoning, separates run management from task work. A controller assesses progress, proposes next steps, weighs each against the remaining model-call budget, then dispatches agents with selected context. A matched direct-control baseline uses the same underlying models and workers but chooses each next action in one call over accumulated history. Every controller and worker call counts against the same nominal allowance; equal call caps do not mean equal token use or runtime. The controller stores full outputs in retrievable memory and carries forward a compact summary.
ProgramBench asks coding agents to rebuild programs from documentation and an executable reference, then measures how their versions fare on hidden behavior tests. Its score is the average per-problem share of hidden tests passed, not the share of projects fully solved. At 1,200 model calls, the GPT-5.5 meta-reasoning setup scored 71.5%, versus 63.7% for matched direct control. The authors also report 58.0% for Codex, run with its native GPT-5.5 configuration under the same nominal allowance; this is a separate coding-tool baseline, not the matched control.
The controller did not win at every budget. With Opus 4.8 on ProgramBench, it scored 56.6% at 400 calls, below direct control’s 62.7%, then reached 67.2% at 1,200 calls versus 65.3% for direct control. The authors warn that staged control can add overhead and that a mistaken progress assessment may cause useful work to be dropped or misdirect later steps. Results from four benchmarks in one preprint do not establish how the method would perform in deployed software maintenance or what it costs in tokens and time.
Sources
- Dahal et al., “Thinking Before Thinking: Scaling Agentic Inference Through Meta-Reasoning”, 29 September 2026; method, Table 1, ProgramBench results and limitations.
- Yang et al., “ProgramBench: Can Language Models Rebuild Programs From Scratch?”; benchmark task definition and hidden-test scoring.
- Figure 1 from Dahal et al. (2026), CC BY 4.0; diagram of the authors' controller framework.