Microsoft Research open-sources the Agensh multi-agent framework, supporting up to 1024 coding agents
Microsoft Research has open-sourced Agensh, a multi-agent framework that removes the central orchestrator, with each worker claiming subtasks on its own, sharing findings, and merging code into the same Git repository. On ProgramBench, the team scaled from 1 agent to 1024, raising the average pass rate on the 5 hardest tasks from 19.31% to 28.78% (with 128 agents), and the single pandoc task reached 79.59% with GPT-6.1-sol paired with 32 agents. The code is released under the MIT license, with a default of 8 workers per single node; the 1024-agent run used 16 nodes with 64 agents per node.
Why read this
SelectedAI score 80
The original text presents a multi-agent collaboration mechanism without a central orchestrator and scaling data from 1 to 1024 agents, making it possible to observe the division of labor that emerges as scale increases.
Topics
Body
Add as a preferredsource on Google
Microsoft Research has open-sourced Agensh, a decentralized harness for large teams of coding agents. The Agensh multi-agent harness removes the central orchestrator that most agent frameworks depend on. Every worker picks its own sub-task, shares findings and merges code into a common Git repo. On ProgramBench, the team scaled the setup from 1 agent to 1,024 agents.
TL;DR
A Harness: Tested from 1 to 1,024 agents. Each worker gets a 272,000-token input and 128,000-token output limit.
Runs on: Linux with Docker Compose. Single node by default (8 workers). The 1,024-agent run used 16 nodes, 64 agents each. Workers have a 64 GiB memory ceiling.
Performance: more agents raised scores on all 5 hardest ProgramBench tasks, and reached them sooner.
Best: 79.59% pandoc pass rate with 32 agents on GPT-6.1-sol. GPT-5.6-sol hit 55.06% with 1,024 agents.
Worst: the 5-task mean reached only 28.78% at 128 agents. Going from 1 to 8 agents added 1.37 points.
What is Agensh?
Agensh is a multi-agent harness from Microsoft Research that coordinates many LLM agents without a lead agent.
TEXTTABLELAYOUTIs your PDF parser actually getting it right?omni_parse_bench · scores text, tables & layout · open sourceStar on GitHub→Sponsored
The target problem is the orchestrator ceiling. In orchestrator-worker designs, a lead agent plans, assigns and merges all work. The research paper cites Claude Code agent teams, Codex sub-agents, Copilot fleet and Kimi Agent Swarm as examples. Agensh lets the workers do that coordination themselves.
How does the Agensh cooperation loop work?
Each worker runs the same 5-step loop, concurrently and asynchronously:
Gather context: read the goal, peer progress, messages and findings.
Claim a sub-task: post a CLAIM, then settle overlaps by direct message.
Take action: work locally with tools and post useful findings early.
Verify results: check the work against the sub-task’s acceptance criteria.
Merge progress: merge into the shared repo, post an update, then repeat.
Three services connect the workers. Gitea hosts the shared workspace, with branches, PRs and merge conflicts. Mattermost carries a team channel plus direct messages for urgent collisions. A shared context board, adapted from DeLM, stores typed notes: OBSERVED, FACT, FAIL, CLAIM and PATCH_SUMMARY.
The loop lives in the prompt, not in runtime code. Every worker gets the same prompt except its ID. Agensh sits on top of a single-agent harness. The experiments used GitHub Copilot CLI. The research paper states thin adapters can connect other harnesses, such as Claude Code.
How well does Agensh scale on ProgramBench?
Microsoft team tested the 5 hardest ProgramBench tasks: FFmpeg, gromacs, pandoc, PHP-src and ctags. Agents rebuild each program from its observed behavior. They work offline, within 6 hours. Reference repos range from 2,767 files (pandoc) to 26,266 files (PHP-src).
With GPT-5.6-sol at high reasoning effort, the mean final test-pass rate moved as follows:
That is a 9.47-point gain, about 49% relative. On pandoc alone, scores were 33.89% (1), 41.75% (8), 43.11% (32), 50.94% (128) and 55.06% (1,024).
Speed improved too: on pandoc, 128 agents passed 30% at the 30-minute checkpoint. 32 and 8 agents needed 60 and 90 minutes.
The README adds a newer model. With GPT-6.1-sol, pandoc rose from 56.19% (1 agent) to 79.59% (32 agents). That is a 41.64% relative gain. These GPT-6.1-sol figures are not in the v1 paper.
What behaviors emerged as the team grew?
The research team traced worker trajectories at each scale:
8 agents: peers agreed on interfaces and resolved overlapping claims themselves.
32 agents: peers reviewed shared work; one approval was withdrawn after a counterexample.
128 agents: workers reused trusted reviewers and standardized an integration protocol.
1,024 agents: several workers acted as integrators, and peers recovered failed work.
None of these roles was hard-coded.
How does Agensh compare with other multi-agent harnesses?
FeatureAgenshClaude Code agent teamsKimi Agent SwarmCopilot CLI /fleet
CoordinationNo orchestrator; peers self-claim workFixed lead session coordinates teammatesOrchestrator decomposes tasksCopilot assigns parts to subagents
Agents shown1,024 testedNo hard limit; 3 to 5 recommendedUp to 100 sub-agentsNot disclosed
Shared stateGitea repo + shared context boardShared task list with file lockingNot disclosedPlan from plan mode
MessagingMattermost channel + direct messagesMailbox messages between teammatesNot disclosedNot disclosed
ModelsModel agnostic; tested GPT-5.6-sol, GPT-6.1-solClaude modelsKimi K2.5Not disclosed
Speed claim30% on pandoc at 30 min (128 agents)Not disclosed4.5x faster than sequential (self-reported)Not disclosed
How do you run Agensh?
The release includes a single-node Docker Compose team and a single-agent baseline. Both use native Copilot ACP sessions. You need Linux with Docker Engine, Compose v2, systemd, cgroup v2 and Python 3.11+. Defaults are 8 workers, 6 hours per task and gpt-5.6-sol at high effort. You also need ProgramBench (v1.2.5 was validated) and a model API token. Full steps sit in the reproduction guide.
Key Takeaways
Agensh coordinates up to 1,024 agents with no central orchestrator.
Workers self-claim tasks via Gitea, Mattermost and a shared findings board.
128 agents lifted the 5-task ProgramBench mean from 19.31% to 28.78%.
GPT-6.1-sol with 32 agents hit 79.59% on pandoc.
Code is MIT licensed, but token cost is not disclosed.
Asif RazzaqWebsite | + postsBio ⮌
Asif RazzaqNVIDIA Releases Nemotron 3.5 Bolt E4B 4GB: An Open Small Language Model for Local Voice Agents, Game Characters and Robots
Asif RazzaqMoonshot AI Releases Kimi K3.1 Pro: A 2.8T Total/104B Active Params (MoE) Model with 1M Context
$uv pip install omni-parse-bench✓text✓tables✓layout→ scored per pageStar on GitHub→Sponsored
RELATED ARTICLESMORE FROM AUTHOR
How to Run Underdog Saluki 27B: 2-Bit Qwen3.8-27B Tool-Calling Guide
Meet Anthropic’s sandbox-runtime: OS-Level Sandboxing for AI Agents, No Containers
NVIDIA Releases Nemotron 3.5 Bolt E4B 4GB: An Open Small Language Model for Local Voice Agents, Game Characters and Robots