Agensh: Decentralized multi-agent harness for 1,024 agents

What changed

Microsoft Research has open-sourced Agensh, a decentralized coordination system for teams of LLM agents. The code became available on October 24, 2026, on GitHub. Unlike traditional multi-agent frameworks that assign a lead agent to plan and manage all work, Agensh allows workers to coordinate themselves through shared infrastructure: a Git repository, a messaging system, and a findings board.

The approach addresses what researchers describe as the orchestrator ceiling. Most multi-agent designs depend on a lead agent to handle task assignments and result merging, similar to patterns in Claude Code agent teams, Copilot fleet, and Kimi Agent Swarm. Agensh eliminates this bottleneck by letting each agent independently choose its sub-task, report findings, and merge progress without central control.

Microsoft tested Agensh on the five most difficult ProgramBench tasks: FFmpeg, gromacs, pandoc, PHP-src, and ctags. Each agent operated within a 272,000-token input budget and 128,000-token output budget. The setup defaulted to 8 workers on a single node. The largest run deployed 1,024 agents across 16 nodes with 64 agents per node. All workers operated within a 64 GiB memory limit.

Why it matters

Agensh solves a practical constraint in multi-agent workflows. As agent teams grow, central orchestration often becomes a performance bottleneck. Agensh demonstrates that decentralized coordination can handle scale and deliver measurable improvements in task performance.

On ProgramBench tasks, adding agents consistently improved results. Using GPT-5.6-sol at high reasoning effort, the mean pass rate across five tasks increased from 19.31% with one agent to 28.78% with 128 agents. On pandoc specifically, the pass rate climbed from 33.89% with one agent to 50.94% with 128 agents and 55.06% with 1,024 agents. With the newer GPT-6.1-sol model, pandoc achieved 79.59% pass rate with just 32 agents, compared to 56.19% with a single agent.

Scaling also improved speed. On pandoc, teams of 128 agents reached 30% pass rate at the 30-minute mark. Teams of 32 agents and 8 agents required 60 and 90 minutes respectively to reach the same threshold.

For organizations building multi-agent systems, the emergence of self-organized behavior is significant. At 8 agents, workers resolved overlapping task claims through direct coordination. At 32 agents, workers reviewed shared code and withdrew approvals when counterexamples appeared. By 128 agents, trusted reviewers became recurring contributors and teams developed standardized integration procedures. At 1,024 agents, some workers specialized as integrators and recovered failed work from peers. None of these coordinating roles required pre-programmed assignments.

What to test

Before implementing Agensh, verify these key aspects and limitations:

• Performance scaling claims: Run the provided ProgramBench experiments to confirm reported gains on your hardware and model choices. The 49% relative improvement on five tasks and the 41.64% improvement on pandoc with GPT-6.1-sol should be reproducible using the included reproduction guide.

• Token cost estimation: The code is MIT licensed but token costs are not disclosed. Calculate total inference costs for your intended agent count and model. At 1,024 agents with 272,000-token inputs per cycle, token consumption accumulates rapidly.

• Infrastructure compatibility: Confirm your Linux environment provides Docker Engine, Docker Compose v2, systemd, cgroup v2, and Python 3.11 or higher. Test the single-node Docker Compose setup before scaling to multiple nodes.

• Model support: Microsoft’s experiments used Copilot ACP sessions with GPT-5.6-sol and GPT-6.1-sol. Test whether Agensh’s thin adapter approach works with alternative language models. The GitHub README indicates adapters could connect Claude Code or other harnesses, but this has not yet been demonstrated.

• Time constraints under real conditions: The tests allowed six hours per task. Benchmark task completion quality when time limits are tighter and determine whether accuracy degrades significantly.

• Coordination reliability at scale: While self-organized roles emerged during testing, observe whether worker conflicts or coordination breakdowns occur at your target team size. The research does not report failure rates or describe scenarios where workers could not resolve conflicts.

• Suitability for your workload: Agensh was evaluated on code reconstruction tasks (FFmpeg, gromacs, pandoc, PHP-src, ctags). Results may not transfer to other domains or task types.

The conclusion

Agensh demonstrates that large agent teams can coordinate effectively without central orchestration. The 49% relative improvement on ProgramBench and the spontaneous emergence of specialized roles at scale suggest the design handles complexity gracefully. The MIT license and open-source release remove barriers to experimentation.

Adoption requires evaluating three key factors. First, token costs are not publicly disclosed, so independent estimation is necessary. Second, infrastructure must meet specific Linux, Docker, and Python requirements, and scaling from 8 agents to 1,024 requires 16 nodes and careful resource allocation. Third, the benchmark tasks involve reconstructing programs from behavior and documentation, which may not match your use cases.

The practical next question is performance in production. The tests ran in a controlled environment: offline tasks, six-hour budgets, no Internet connectivity. Monitor for published case studies showing Agensh performance on production workflows that include interruptions, network latency, and tasks outside ProgramBench scope. Also track whether other multi-agent frameworks adopt similar decentralized patterns or whether Agensh remains specialized for large coding teams.

For teams already using agent security frameworks like Anthropic’s sandbox-runtime, Agensh serves a different purpose: scaling task coordination rather than enforcing per-agent isolation. Consider it as a supplementary layer above per-agent security measures, not as a replacement. Begin with the Docker Compose single-node setup and follow the reproduction guide to understand the operational requirements and capabilities.

Related reading

• Google Agent Compute Engine: microVM isolation, zero credentials
• Grok Voice 3.0 vs Voxtral 3: Voice Agent Comparison