AI
Self-Evolving Harness on Multiple Tasks with the Agent as Its Own Optimizer
Publication: cs.AI updates on arXiv.org
Author: Qiankai Xu
Source published: 01/10/2026, 04:00:00 UTC
Briefing created: 01/10/2026, 04:09:00 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 01/10/2026, 04:03:24 UTC.
Briefings
A new framework enables agents to evolve their own execution harnesses through recursive self-improvement. Starting from a 49-line seed harness, the system solves tasks on five benchmarks, then uses those run records to directly edit the harness. This process separates in-distribution and out-of-distribution tasks to measure generalization. The first stage improved average scores by 4.48 points on in-distribution benchmarks and 12.64 points on out-of-distribution benchmarks, surpassing Codex on the former while matching it on the latter. Continued evolution on the Claw-Eval benchmark raised scores from 66.17 to 68.06, exceeding Codex. The authors analyzed emergent mechanisms like output truncation and history compaction.
Key points
- The framework uses a frozen model and a single version of the harness to solve tasks, then edits the harness directly based on run records.
- Evolution draws tasks from five benchmarks across different domains, with strict separation between training and held-out tasks to measure generalization.
Context and limitations
Evaluation is limited to five benchmarks across different domains, with only two out-of-distribution benchmarks used for final evaluation. The study does not report results on all five out-of-distribution benchmarks.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 01/10/2026, 04:09:00 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeif74sellfu4wfdisatrwmbuuvp7roicglsuddjlwkf7htno3z3cey