Microsoft's Agensh scales to 1,024 agents, lifting pandoc pass rate to 55%
MicrosoftResearch · hf · 2026-09-23
Microsoft Research introduced Agensh, a self-organized multi-agent harness with no central orchestrator, exploring agent count as a new scaling dimension.
- Motivation: existing multi-agent harnesses are often bottlenecked by a central orchestrator's capacity to allocate tasks. Agensh instead runs a cooperation loop where concurrent workers gather context, self-assign subtasks, act, share findings, verify, and merge progress asynchronously.
- Infrastructure: three components—a shared workspace for proposed/ongoing/completed work, a message interface for worker communication, and shared context retaining reusable findings and intentions.
- Results: on the five hardest ProgramBench tasks with GPT-5.6-sol (high), scaling from 1 to 128 agents raised mean test-pass rate from 19.31% to 28.78% (49% relative improvement); larger organizations reach comparable rates earlier.
- On pandoc, scaling from 1 to 1,024 agents lifted the pass rate from 33.89% to 55.06%. Worker trajectories show self-organized cooperation forms emerging and standardizing as the organization grows.
Related event: Microsoft Unveils Agensh, a Self-Organized Multi-Agent Framework(2 posts)→
More from coding & agent
- The prompt to run before wiring an agent to another service — gethackteam · 2026-09-23
- Vercel CEO Backs px0, a Lightweight IDE Built for Reviewing Agent-Written Code — arpit_bhayani · 2026-09-23
- Matt Pocock: Stop chasing model releases, improve your agent's harness instead — mattpocockuk · 2026-09-23
- Uncle Bob: AI changes nothing—complexity, not tooling, still makes software slow — blaizedsouza · 2026-09-23
- GBrain: plug your own memory, tools, and skills into any AI — garrytan · 2026-09-23
- Podcast: building a playable game with $8 of parts and AI assistance — aishashok14 · 2026-09-23