Microsoft Research's Agensh scales coding agents to 1,024 without an orchestrator

An arXiv preprint reports test-pass gains from 19.31% to 28.78% as an orchestrator-free agent harness scales from 1 to 128 agents.

Microsoft Research has posted Agensh, a multi-agent coding harness with no central orchestrator. It is a preprint. The authors report that adding agents improves results. A critique argues the gain is mostly extra compute and noisier than the headline suggests, which is a useful check for anyone whose plan is simply to add more subagents.

In Agensh, each worker runs the same loop asynchronously. It gathers context, claims a sub-task, does the work, shares findings, verifies the result and merges it. Coordination happens through three shared pieces: a workspace for proposed, ongoing and completed work, a message interface, and a shared context of findings and intentions. AlphaSignal describes these as a Git workspace, a Mattermost message bus and a typed context board.

The paper's results use GPT-5.6-sol (high) on the five hardest ProgramBench tasks:

The authors also say larger teams reach comparable pass rates sooner.

There are caveats. A critique on GitHub makes several points:

Whether the results come from single runs or include run-to-run variance isn't confirmed. It's also unclear whether the code is public. A proposal in ruvnet/metaharness suggests re-testing against best-of-N or longer single-agent rollouts at equal total compute. The paper doesn't include that baseline, and no head-to-head comparison with orchestrator-based harnesses was found. For now, read Agensh as a way to trade compute for latency, not proof that coordination is more efficient.

Sources