Microsoft Open-Sources Agensh: A Decentralized Multi-Agent Harness That Scales to 1,024 Coding Agents
Microsoft Research has open-sourced Agensh, a decentralized harness for large teams of coding agents. Unlike most agent frameworks, Agensh removes the central orchestrator entirely. Each worker selects its own sub-task, shares findings with its peers, and merges code into a common Git repository. On ProgramBench, the team scaled the setup from a single agent to 1,024 agents — a meaningful milestone for 2026's agentic-AI infrastructure landscape, where coordination overhead, not raw model capability, has become the primary scaling bottleneck.
TL;DR
- Harness scope: Tested from 1 to 1,024 agents. Each worker receives a 272,000-token input limit and a 128,000-token output limit.
- Runtime: Linux with Docker Compose. A single node runs by default (8 workers). The 1,024-agent run used 16 nodes with 64 agents each. Workers are capped at 64 GiB of memory.
- Performance: More agents raised scores on all five of the hardest ProgramBench tasks — and reached those scores sooner.
- Best result: A 79.59% pass rate on pandoc with 32 agents running GPT-6.1-sol. GPT-5.6-sol reached 55.06% with 1,024 agents.
- Worst result: The five-task mean reached only 28.78% at 128 agents. Scaling from 1 to 8 agents added just 1.37 points.
What Is Agensh?
Agensh is a multi-agent harness from Microsoft Research that coordinates many LLM agents without a lead agent. Rather than routing every decision through a central planner, workers operate autonomously against a shared substrate: a common Git repository and a shared pool of findings. This design reflects a broader 2026 shift toward decentralized, peer-to-peer agent coordination, as teams push past the orchestration limits that constrain hub-and-spoke frameworks at high agent counts.
Article originally reported by Asif Razzaq for Marktechpost on October 24, 2026.
via MarkTechPost