A harness without an orchestrator
Microsoft Research has open-sourced Agensh, a harness for running large teams of coding agents. Unlike most agent frameworks, there is no lead agent that plans and assigns work. Each worker claims its own sub-task, shares findings and merges code into a shared Git repository.
How agents coordinate
Every worker runs the same five-step loop concurrently and asynchronously: gather context, claim a sub-task, take action, verify results and merge progress. Coordination runs through three services: Gitea for the shared workspace, Mattermost for announcements and urgent collisions, and a shared context board for short typed notes. The loop lives in the prompt rather than in runtime code.
Scaling results
The team tested the five hardest ProgramBench tasks, scaling from 1 to 1,024 agents. More agents raised scores and reached targets sooner. The best result was a 79.59% pandoc pass rate with 32 agents on GPT-6.1-sol. With GPT-5.6-sol, 1,024 agents reached 55.06%. The five-task mean stayed at just 28.78% with 128 agents.
Caveats
Experiments ran with a six-hour budget, no internet access and high reasoning effort, using GitHub Copilot CLI as the single-agent harness. The paper says thin adapters could connect other harnesses such as Claude Code. Only pandoc ran at 1,024 agents, and token cost was not disclosed.
