Microsoft's Agensh Runs 1,024 Coding Agents Without a Central Controller

Microsoft Research’s new Agensh system coordinated as many as 1,024 coding agents working asynchronously on a shared codebase. Most multi-agent coding systems rely on a central agent to divide work, supervise workers, and combine their output. Agensh distributes those duties across shared services and a common worker protocol, removing the coordinator’s context window as the main scaling constraint.

One context window, one ceiling

Codex subagents, Claude Code subagents and agent teams, Copilot fleet, and Kimi Agent Swarm use an orchestrator-worker structure. A central agent plans the job, assigns subtasks, tracks progress, handles conflicts, and integrates contributions. As the worker pool grows, that agent must fit more workstreams, messages, and dependencies into one context window.

Agensh moves coordination into three shared components that every worker can access:

• Workspace: A shared repository contains the organization’s current and integrated work.
• Messages: Organization-wide announcements and direct messages carry updates that require attention.
• Shared context: A structured board records reusable findings, failed approaches, active claims, and work intentions.

Workers consult those components independently, choose work, communicate with peers, and merge completed changes. They run concurrently and asynchronously, so progress does not depend on a central queue or a global synchronization step.

Five steps per worker

Each worker repeatedly executes the same five-step loop:

1. Gather context from the repository, messages, and shared context board to identify unfinished work.
2. Claim a subtask by adding a CLAIM record, then resolve overlaps through direct messages.
3. Take action with local tools and publish useful intermediate findings.