Self-Harness
Free while signed in. Answers cite the passages they came from.
Most agent scaffolds are built once by hand and then frozen, even as the underlying models keep changing. This paper introduces Self-Harness, a paradigm where an LLM agent improves its own operating harness, the prompts, tools, memory, and orchestration around the base model, without human engineers or a stronger external agent. Because every model fails in its own way, the system mines those model-specific weaknesses and turns them into concrete, executable harness edits rather than generic advice.
A three-stage self-improvement loop: Self-Harness runs Weakness Mining, which clusters execution traces into model-specific failure patterns, then Harness Proposal, which generates diverse but minimal edits tied to those failures, then Proposal Validation, which accepts edits only after regression testing on held-in and held-out splits.
Consistent gains across base models: On Terminal-Bench-2.0, held-out pass rates rise for every model tested. MiniMax M2.5 improves from 40.5% to 61.9%, Qwen3.5-35B-A3B from 23.8% to 38.1%, and GLM-5 from 42.9% to 57.1%.
Weaknesses become edits: Rather than appending generic instructions, the loop converts each observed failure mode into a targeted change to memory, tools, or prompts, with reported relative improvements as high as 138%.
Why it matters: As models proliferate and evolve, hand-tuning a bespoke harness for each one does not scale. Self-Harness shows the scaffold itself can be made to adapt, closing the gap between a frozen harness and the model it wraps.
Get next week’s papers.
The same picks and the same summaries, in your inbox. Free, and 176 issues deep.
Subscribe on Substack