Microsoft Open-Sources Agensh: How Its 1,024-Agent Harness Works

Microsoft Research has open-sourced Agensh, a harness for coding agents that coordinate without a central manager. The October 24 release includes the infrastructure and experiment scripts behind research involving up to 1,024 agents, giving developers a way to inspect how large agent teams divide work and combine results.

The repository announcement separates this code release from the paper published on September 22. That distinction matters: the new development is access to an implementation and reproduction guide, rather than a newly announced model or a hosted service anyone can switch on.

How Agensh coordinates agents

A harness is the software around a model that manages its ongoing work. Agensh adds a cooperation layer above each worker's individual agent loop. In the released setup, workers use native Copilot ACP sessions, while the organization shares a repository, messages and reusable context.

Each worker looks for useful work, claims a task, acts, checks results and contributes back. Other workers can keep moving while that happens. Gitea provides the shared code workspace and Mattermost handles communication; there is no single model session responsible for assigning every next step.

Consider agents rebuilding a document converter. One might examine input handling while another investigates output formatting. The difficult part comes when those implementations need the same interface or change overlapping files. A useful team has to negotiate those dependencies and integrate the result, rather than simply produce several plausible patches.

The interesting question is how much parallel work survives integration. Creating more workers is comparatively straightforward. Keeping their findings, code and assumptions consistent is where a shared harness earns its place.

What the 1,024-agent result measures

The research paper evaluates reconstruction of existing programs from observed behavior and supplied documentation, with internet access disabled. Its main comparison uses GPT-5.6-sol at high reasoning effort and a six-hour budget across five difficult ProgramBench tasks.

Experiment	One agent	Larger team
Mean test-pass rate across five tasks	19.31%	28.78% with 128 agents
Pandoc test-pass rate	33.89%	50.94% with 128 agents
Pandoc test-pass rate	33.89%	55.06% with 1,024 agents

The five-task gain is 9.47 percentage points, approximately 49% relative. The thousand-agent result covers Pandoc specifically, rather than every task. Even that largest team still fails a substantial share of tests.

These are author-reported experiments under a common time limit. They do not establish that adding agents is cheaper, that output improves proportionally with worker count, or that the same gains carry over to everyday repository maintenance. The useful comparison for a product team would also include total model spend, infrastructure load and human review time.

Shared memory needs evidence checks

Agensh's shared-context service retains short typed entries for observations, facts, failed approaches, work claims and patch summaries. Recent entries support prompt delivery, while persisted history remains searchable when older information becomes relevant.

The service also expires live work claims and handles replacement or invalidation of certain patch summaries. That helps a team avoid relying indefinitely on an abandoned assignment or an earlier candidate that a worker has subsequently reported as failed.

There is an important limit: its checks are deterministic text checks. The service does not execute the verification commands an agent says it ran. A shared entry can make a claim visible to everyone without independently proving the claim correct.

For the document-converter example, a note saying that a parser passes a test is useful coordination data. It still needs to match an actual test result and the version of the parser eventually merged. A memory system can distribute a mistake just as efficiently as a useful finding.

What the public setup actually requires

The reproduction guide describes a single-node team and a separate single-agent baseline. Defaults are eight workers, six hours, GPT-5.6-sol and high reasoning effort. Linux, Docker Engine, Compose, systemd, cgroup v2, ProgramBench and a model gateway are prerequisites.

Larger configurations require manual attention. Mattermost's default user limit and MySQL connections do not automatically adjust with the worker count. Startup is staggered, and that startup time counts against the team's budget, so later workers can have less active time.

The guide also documents differences from the historical experiments, including updated cleanroom images and lifecycle behavior. Teams export their final commit before shutdown, while the standalone baseline can retain work produced during cancellation. A fresh run should record its own configuration and revision rather than be presented as an exact recreation of the published result.

The deployment documentation makes the infrastructure concrete: each agent has a worker/router pair, shared services hold team state, and network separation controls model egress. This is an experimental system to operate, with resources and credentials to manage, even though its code uses an MIT license.

Where this release is useful

Agensh gives researchers and agent-infrastructure developers something more useful than a claim that large teams work: an implementation whose coordination decisions can be examined and changed. It is especially relevant when completion time matters enough to justify parallel compute.

For teams building coding products, the next question is narrower than whether a thousand agents can collaborate. It is whether a manageable number can produce more verified, integrated work within the team's actual budget. The open release makes that question testable without treating the biggest experiment as the default deployment plan.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn agent architectures, research releases and product updates into clear stories that developers can understand and share.

Book a 30-minute call.