Microsoft Open-Sources Agensh, a Harness That Coordinates 1,024 Coding Agents With No Central Orchestrator

Multi-agent coding systems usually have a boss: a central orchestrator that decides which worker does what, routes messages, and stitches the results together. Microsoft Research has published and open-sourced a harness called Agensh that removes that boss entirely. Coding agents in Agensh organize themselves, and the repo shows the approach scaling to 1,024 simultaneous agents working on a single software task.

The design is deliberately flat. Every worker runs the same loop asynchronously: gather context, find and claim an unclaimed subtask, act on it, share what was learned, verify the result, and merge progress. Three pieces of shared infrastructure hold the swarm together — a shared workspace that tracks proposed, ongoing and completed work; a message interface that lets workers talk to each other; and a shared context that retains reusable findings and intentions so later workers do not rediscover the same facts.

The headline numbers come from the five hardest tasks in ProgramBench, where agents are asked to reconstruct a reference program purely by observing its behavior, with no internet access and a six-hour budget, using GPT-5.6-sol at high reasoning effort. Going from one agent to 128 lifted the mean final test-pass rate from 19.31 percent to 28.78 percent, a relative improvement of about 49 percent, with larger teams reaching a given pass rate sooner. On the single hardest task, building the pandoc document converter from scratch, pass rates climbed from 33.89 percent with one agent to 50.94 percent with 128 and 55.06 percent with 1,024. Swapping in GPT-6.1-sol pushed pandoc from 56.19 percent to 79.59 percent at just 32 agents, a 41.6 percent relative gain.

The paper also documents something less quantitative: as the team grows, the agents start inventing their own division of labor. Peer-coordination habits, integration procedures and de-facto specialist roles emerged from the workers themselves and hardened into standard practice at larger scales. The authors frame the number of agents as a new scaling axis — one that matters most for complex tasks running under hard latency limits or fixed time budgets, where waiting for a single long trajectory is the bottleneck.

Every performance figure above comes from Microsoft's own runs on its own benchmark selection, and no independent replication has been published. The gains are also relative to a very low baseline: a 19.31 percent solo pass rate means the five chosen tasks are near the frontier of what a single agent can do, which is exactly where orchestration overhead tends to matter most. Token cost, a sore point in harness research this month, is reported only indirectly — larger teams finish sooner, which implies fewer wall-clock hours but not necessarily fewer tokens.

The release lands in a noisy week for harness design. Anthropic published guidance arguing that most multi-tier subagent structures should be folded back into a single loop, and a Shandong University study released Friday found that swapping coding-agent harnesses barely moves pass rates while sometimes tripling the bill. Agensh bets the opposite way: instead of fewer agents, more of them — as long as nobody is waiting for a coordinator. Both positions can be true at once; they optimize different bottlenecks.

The repo is on GitHub under Microsoft's research organization, with code, prompts and evaluation setup available for teams that want to test the claims on their own workloads. If the self-organization result holds up outside Redmond, the practical consequence is modest but real: time-boxed, hard tasks — long rebuilds, large migrations, exhaustive verification — get a new dial to turn. For everything else, a single agent with good tools remains the cheaper default.