Microsoft open-sources Agensh, a lead-free multi-agent harness scaling to 1,024 coding agents
solyarisoftware · x · 2026-10-26
Microsoft Research open-sourced Agensh, a decentralized multi-agent harness that removes the central orchestrator and scales to 1,024 coding agents.
How it works: self-organization instead of orchestration. Each worker runs the same loop — claim a sub-task, build and verify locally, merge via PR into a shared Gitea repo — coordinating through the shared repo, a Mattermost channel and a findings board. Integrator and reviewer roles emerged on their own at 128+ agents.
Benchmarks (5 hardest ProgramBench tasks, 6-hour offline runs):
- GPT-5.6-sol: mean test-pass rate rose from 19.31% (1 agent) to 28.78% (128 agents)
- pandoc: 1,024 agents hit 55.06% (vs 33.89% with 1); GPT-6.1-sol reached 79.59% with just 32 agents
- Diminishing returns: 1→8 agents added only 1.37 pts; token cost per run undisclosed
Deployment: MIT-licensed, Linux + Docker Compose; default single node with 8 workers, the 1,024-agent run used 16 nodes (64 agents each, 64 GiB memory ceiling, 272K input / 128K output token limits per worker).
More from coding & agent
- 'Most people run Claude at 30%': the surgical revision prompt that beats 'make it better' — JafarNajafov · 2026-10-26
- Reddit asks: Which AI engineering resources actually helped you ship value this year? — nian2326076 · 2026-10-26
- Agent procrastination debate: repligate argues AI does better work when allowed to pace itself — repligate · 2026-10-26
- MCP Python SDK package naming confusion trips up server developers — Mr_Unknown_Hero · 2026-10-26
- The math on shipping tiny Shopify apps with Claude Code: one app, up to $8,740/month — thisdudelikesAI · 2026-10-26
- Background CUA is the future: separate execution environments and handoff matter for computer-use agents — TianbaoX · 2026-10-26