Microsoft Research open-sources Agensh: 1,024 coding agents, no boss, 55% on Pandoc rebuild
JafarNajafov · x · 2026-10-31
Microsoft Research open-sourced Agensh, a decentralized multi-agent coding harness that scales to 1,024 agents with no central orchestrator or lead. Each agent gets near-identical instructions and loops through: gather context, claim a sub-task, act, verify, merge; a shared workspace, message interface and shared context coordinate the swarm.
On the five hardest ProgramBench tasks (GPT-5.6-sol, 6-hour budget, no internet), mean test-pass rate rose from 19.31% with 1 agent to 28.78% with 128 (49% relative). On rebuilding Pandoc, pass rate went 33.89% → 50.94% (128 agents) → 55.06% (1,024 agents). With GPT-6.1-sol, Pandoc rose from 56.19% to 79.59% with 32 agents (+41.64% relative). Peer coordination, integration workflows and specialized roles emerge as teams grow. Copilot, Codex, Claude Code and Pi are supported as single-agent backends.
More from coding & agent
- Uber uses AI agents to delete 1M+ lines of code and 16,000 stale feature flags — bibryam · 2026-10-31
- Tool builder details agent-first Apple platform dev: CarPlay sims, build queueing, App Store Connect automation — rxwei · 2026-10-31
- Indie dev runs Dark Veil's all-platform releases entirely on Codex agents — Dimillian · 2026-10-31
- 10 Open-Source MCP Projects to Give AI Agents Web Search and Crawling — Shruti_0810 · 2026-10-31
- Claude iterated on its own screenshots against 341 real photos to build a browser island — prasenx · 2026-10-31
- Filmmaker used OpenClaw to organize notes and win a grant: workflow beats tools — bytebot · 2026-10-31