Agensh: Scaling Organizational Intelligence to 1,024 Agents

A multi-agent coding harness from Microsoft Research's General AI group (Furu Wei, Li Dong, Yan Xia corresponding; Zhihao Zhan and Ting Song first authors) that drops the central orchestrator. Every worker runs the same asynchronous cooperation loop from the same prompt, differing only by worker ID: gather context and read peer progress, claim a sub-task, resolve overlaps by message, act, publish findings, verify, merge. Three pieces of agentic organization infrastructure hold it together: a shared workspace (a Gitea instance, with private branches merged to main and conflicts surfaced for workers to resolve), a message interface (Mattermost, with a task channel delivered at loop start and direct messages delivered after each tool call), and a typed shared context of OBSERVED, FACT, FAIL and CLAIM entries adapted from DeLM. Agensh sits above an existing single-agent harness through a thin adapter; the paper's runs use GitHub Copilot as the inner loop, with Claude Code named as another target.

The test is the five hardest of ProgramBench's 200 tasks (FFmpeg, gromacs, pandoc, PHP-src, ctags: rebuild the reference software from scratch in six hours with no Internet), all with GPT-5.6 Sol at high effort. Mean final test-pass rate rises from 19.31% with one agent to 20.68%, 26.52% and 28.78% with 8, 32 and 128 agents, a 49% relative gain, and larger organizations get there sooner: on pandoc, 128 agents pass 30% at the 30-minute mark, 32 agents at 60 and 8 agents at 90. Pushing pandoc to 1,024 agents lifts the final rate from 33.89% to 50.94% at 128 and 55.06% at 1,024. Trajectories show cooperation forms that appear only as the organization grows, from interface agreements between peers to multi-worker integration, standardized workflows and specialized roles. The authors position agent count as a scaling dimension of its own, distinct from orchestrator-worker designs such as Claude Code agent teams, Codex subagents, Copilot fleet and Kimi Agent Swarm; a sibling approach from Shanghai AI Lab is Harness-of-Harness. The paper lists a code release at github.com/microsoft/Agensh, which did not exist at filing. No cost or token accounting is reported.