Meta paper: a dedicated controller lifts long agent runs from 63.7% to 71.5% at same budget
dair_ai · x · 2026-10-02
Meta Superintelligence Labs published a paper on controlling long agent runs, showing that a dedicated controller deciding what work to run next significantly outperforms direct control.
- With the same workers and budget, GPT-5.5 on ProgramBench improves from 63.7% to 71.5% (+7.8 points); Codex scores 58.0%.
- The controller keeps a short run summary while full worker outputs stay in memory; each cycle it updates the summary, proposes next steps, estimates each option's value under the remaining budget, and routes chosen work—with earlier outputs—to workers.
- Across ProofBench, ARC-AGI-2 and LongCoT-mini, it adds 3.6–4.2 points on average over direct control, tested on three frontier models.
Paper: https://t.co/34PuaVLnkr
More from coding & agent
- AI in finance shows a power law: top 1% users generate 35x more files with agents — peterjliu · 2026-11-19
- Linux LTS patch flood: 3318 patches for 6.6 alone, 'no way humans are reviewing this' — pkqzy888 · 2026-11-19
- Caliper Harness: open-source coding harness disciplines small local models with plans, checks, reviews — Saraozte01 · 2026-11-19
- AI systems lack executive function, not scale: four manager-layer strategies to fix it — Designer-Material633 · 2026-11-19
- Google employee replaces doomscrolling with 3 fully automated personal AI newsletters — DynamicWebPaige · 2026-11-19
- Designer spins up an agent to build state-airline brand gallery mid-flight delay — beh_zod · 2026-11-19