The Hardest Part of Improving an Agent Is Finding the Line to Change
Harness Handbook picked up 163 upvotes on HuggingFace papers, out of Tencent Hunyuan. It goes after a problem nobody writes papers about because it sounds like a chore: when you want to change how your agent behaves, where in the codebase do you actually go?
Start from the premise the paper opens with. An agent's capability isn't just its model, it's the harness around it, the thing managing prompts, state, tools, and execution. Everyone in this field has internalized that by now. What nobody talks about is that harnesses turn into sprawl fast. The behavior you want to change is spread across four files, one of them a code path that runs twice a week, and the model you asked to fix it confidently edits the wrong one.
The Handbook is a behavior-centric map of the harness, generated automatically from the codebase with static analysis plus LLM-assisted structuring. Pair it with Behavior-Guided Progressive Disclosure, which walks you down from a high-level behavior to the specific implementation, and the agent stops guessing. On open-source harnesses, Handbook-assisted planning improved both behavior localization and edit quality while using fewer tokens. Better and cheaper is rare enough to notice.
The wins concentrated exactly where you'd hope: scattered sites, rarely executed paths, and cross-module interactions. Those are the three places a coding agent reliably falls on its face, because none of them are visible from any single file.
Here's why this is more than tooling. Self-improving agents are the whole endgame, and self-improvement means an agent editing its own harness. Every paper in that thread assumes the agent can find the code it needs to change. Harness Handbook is the first one to treat that assumption as the actual bottleneck. An agent that can't navigate itself can't rewrite itself.
https://ruhan-wang.github.io/Harness-Handbook/
← Back to all articles
Start from the premise the paper opens with. An agent's capability isn't just its model, it's the harness around it, the thing managing prompts, state, tools, and execution. Everyone in this field has internalized that by now. What nobody talks about is that harnesses turn into sprawl fast. The behavior you want to change is spread across four files, one of them a code path that runs twice a week, and the model you asked to fix it confidently edits the wrong one.
The Handbook is a behavior-centric map of the harness, generated automatically from the codebase with static analysis plus LLM-assisted structuring. Pair it with Behavior-Guided Progressive Disclosure, which walks you down from a high-level behavior to the specific implementation, and the agent stops guessing. On open-source harnesses, Handbook-assisted planning improved both behavior localization and edit quality while using fewer tokens. Better and cheaper is rare enough to notice.
The wins concentrated exactly where you'd hope: scattered sites, rarely executed paths, and cross-module interactions. Those are the three places a coding agent reliably falls on its face, because none of them are visible from any single file.
Here's why this is more than tooling. Self-improving agents are the whole endgame, and self-improvement means an agent editing its own harness. Every paper in that thread assumes the agent can find the code it needs to change. Harness Handbook is the first one to treat that assumption as the actual bottleneck. An agent that can't navigate itself can't rewrite itself.
https://ruhan-wang.github.io/Harness-Handbook/
Comments