I started using this tool called StateM for my agents lately. It is an open-source control layer for long-horizon agents, and I felt it was worth sharing here.
The basic idea is to move parts of agent control from the prompt or agent context to an externalized layer, making them explicit and enforceable. The interesting thing is what they call an “agent owning the control”: StateM uses YAML-defined runbooks and a CLI runtime to provide things like stateful checks, phase-specific context, blocking transitions, and persistent workflow state. It’s very easy to implement - my agent is already using it and adapting its own runbook when it finds anything that should be updated. I’ve even seen it switch runbooks when I ask it for something off-topic.
I started trying it after seeing how much performance it squeezed out of flagship models in their experiments. They achieved 95.3% on Terminal-Bench 2.1 using GPT-5.6 Sol xhigh + Adapted StateM (+ 10.4 points). It is even higher than GPT-5.6 Sol Ultra’s 91.9%. And DeepSeek-V4-Flash can achieve the same performance as GPT-5.6 Sol Max, using 1/38 of the cost. These numbers are crazy. Is scaling the harness also an orthogonal axis to model scaling?
After seeing these results, I would love to hear what others are thinking. How important should agent harness be in agent reliability? In the long horizon tasks it seems more important than I thought.
I started using this tool called StateM for my agents lately. It is an open-source control layer for long-horizon agents, and I felt it was worth sharing here.
The basic idea is to move parts of agent control from the prompt or agent context to an externalized layer, making them explicit and enforceable. The interesting thing is what they call an “agent owning the control”: StateM uses YAML-defined runbooks and a CLI runtime to provide things like stateful checks, phase-specific context, blocking transitions, and persistent workflow state. It’s very easy to implement - my agent is already using it and adapting its own runbook when it finds anything that should be updated. I’ve even seen it switch runbooks when I ask it for something off-topic.
I started trying it after seeing how much performance it squeezed out of flagship models in their experiments. They achieved 95.3% on Terminal-Bench 2.1 using GPT-5.6 Sol xhigh + Adapted StateM (+ 10.4 points). It is even higher than GPT-5.6 Sol Ultra’s 91.9%. And DeepSeek-V4-Flash can achieve the same performance as GPT-5.6 Sol Max, using 1/38 of the cost. These numbers are crazy. Is scaling the harness also an orthogonal axis to model scaling?
After seeing these results, I would love to hear what others are thinking. How important should agent harness be in agent reliability? In the long horizon tasks it seems more important than I thought.
Github link: https://github.com/henryqin1997/statem Paper: https://arxiv.org/abs/2608.15089
Thanks for sharing! Author here, plz ask me anything!