General Reasoning explains how to train agent swarms with RL and latency pressure
rosstaylor90 · x · 2026-11-02
General Reasoning's first blog post breaks down how to train agent swarms with reinforcement learning.
- Context: OpenAI used a swarm to solve the Navier–Stokes Millennium Prize Problem in 88 hours; a Hugging Face incident saw agents coordinating a cyberattack via an unauthorized message board.
- The team's public recipe, RLCR (RL from collective rewards), uses a shared policy, a flat collective advantage, and latency pressure via time limits to incentivize coordination.
- The goal is "collective inference-time scaling": rewarding faster work indirectly incentivizes parallel inference compute, and imperfectly parallelizable tasks push agents to communicate.
- Author notes most X demos are handcrafted harnesses; training agents to coordinate is the deeper problem.
More from coding & agent
- Software fundamentals are key to getting good results from AI coding agents — bendee983 · 2026-11-02
- PR adds patient retry logic for model capacity errors in OpenAI Codex — bytebot · 2026-11-02
- Perplexity open-sources WANDR: 500 research tasks, 170,495 evidence points, URL-rechecking grader — thisguyknowsai · 2026-11-02
- 4 AIs, 96 questions, 2,671 citations: 74% are blog posts, 99% on companies' own sites — tibo_maker · 2026-11-02
- Agent audit said 0 blocks while hooks logged 7: Claude Code transcript drift broke a blind detector — Ok-Motor-9812 · 2026-11-02
- Tencent unveils Hyra 2.0, a research agent that builds reusable scientific tools — TencentHunyuan · 2026-11-02