- Author
- Hangfan Zhang and seven co-authors, Shanghai Artificial Intelligence Laboratory
- Published
- 20 August 2026 (version 3)
- Kind
- arXiv preprint with benchmark experiments
- Original
- arxiv.org/abs/2606.09498
- Subject
- Harnesses and agent loops
What it is
A team at a Shanghai research lab wrote the preprint. It argues that agent harnesses are hand-built and scale poorly across models. It proposes letting the model propose bounded edits to its own harness, then testing them.
What it says
- Self-Harness lets an agent improve its own harness without human engineers or a stronger agent.
- The loop mines failure patterns, proposes minimal edits, and validates them with regression tests.
- An edit is accepted only if it improves one task split without hurting the other.
- Across nine model and benchmark pairs, every final harness beat the initial one.
- The largest absolute gain was GLM-5 on AppWorld, from 44.4% to 85.0%.
- For MiniMax M2.5, kept edits create output files earlier and stop unproductive tool loops.
- The authors list limits: bounded edits, possible benchmark-specific fixes, and reliance on verifier quality.
What evidence it gives
Table 1 gives held-in, held-out and overall pass rates for three models on three benchmarks. The paper adds evolution charts, code diffs of kept edits and before-and-after trace cases. Each candidate was evaluated with two repeated attempts.
Interests and limits
The authors are researchers who built and tested their own method. They use small fixed subsets, such as 64 Terminal-Bench-2.0 tasks and 100 SWE-bench cases. The held-out split is hidden from the proposer but is used by the promotion gate. The starting harness is minimal, so gains begin from a low base. The results do not compare against human-built harnesses. Only three models were tested.