Microsoft's Agensh Scales Multi-Agent Systems to 1,024 Agents Without a Central Orchestrator, Boosting Test-Pass Rate to 55%
andrew_n_carr · x · 2026-09-23
Microsoft Research (Furu Wei's team) released Agensh, a self-organized multi-agent harness that removes the central orchestrator bottleneck limiting today's frameworks.
Key design: concurrent workers run a cooperation loop — gathering context, self-claiming sub-tasks, acting, sharing findings, verifying, and merging progress asynchronously — backed by a shared workspace, a message interface, and shared context for reusable findings.
Benchmarks (GPT-5.6-sol high on the five hardest ProgramBench tasks):
- Scaling 1 → 128 agents raises mean test-pass rate from 19.31% to 28.78% (49% relative gain)
- On pandoc, scaling 1 → 1,024 agents lifts the pass rate from 33.89% to 55.06%
- Larger organizations reach comparable pass rates earlier
More from coding & agent
- Security audit: autonomous research program XBOW credited with ~12 upstream bug fixes — moyix · 2026-09-23
- Telling agents to use formal verification helps them write better code — sh_reya · 2026-09-23
- Team uses Typesafe's Jev to prompt follow-up questions during ticket creation — TheMoonMidas · 2026-09-23
- The most common evals mistake: skipping error discovery and measuring the wrong thing — FinanceYF5 · 2026-09-23
- AI products are easy to change and hard to predict: evals turn 'good' into repeatable tests — FinanceYF5 · 2026-09-23
- Is MCP a Bad Idea? New Video Questions the Agent Protocol Standard — RelevantEmergency707 · 2026-09-23