{"paper":{"title":"Agentic Harness Engineering: Observability-Driven Automatic Evolution of Coding-Agent Harnesses","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Three observability pillars let coding-agent harnesses evolve autonomously to beat human designs and transfer across benchmarks.","cross_cats":["cs.SE"],"primary_cat":"cs.CL","authors_text":"Chengjun Pan, Hang Yan, Jiahang Lin, Lizhi Lin, Shichun Liu, Shihan Dou, Tao Gui, Xuanjing Huang, Yu-Gang Jiang, Zhenhua Han, Zhiheng Xi","submitted_at":"2026-04-28T16:55:02Z","abstract_excerpt":"Harnesses are now central to coding-agent performance, mediating how models interact with tools and execution environments. Yet harness engineering remains a manual craft, because automating it faces a heterogeneous action space across editable components, voluminous trajectories that bury actionable signal, and edits whose effect is hard to attribute. We introduce Agentic Harness Engineering (AHE), a closed loop that addresses these challenges through three matched observability pillars: (1) component observability gives every editable harness component a file-level representation so the acti"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"ten AHE iterations lift pass@1 on Terminal-Bench 2 from 69.7% to 77.0%, surpassing the human-designed harness Codex-CLI (71.9%) and the self-evolving baselines ACE and TF-GRPO. The frozen harness transfers without re-evolution: on SWE-bench-verified it tops aggregate success at 12% fewer tokens than the seed, and on Terminal-Bench 2 it yields +5.1 to +10.1pp cross-family gains across three alternate model families.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the three observability pillars sufficiently constrain the action space and provide actionable signal so that the evolution loop produces generalizable improvements rather than benchmark-specific overfitting or noise-driven changes.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"AHE automates coding-agent harness evolution via component, experience, and decision observability, raising Terminal-Bench 2 pass@1 from 69.7% to 77.0% with transfer gains across models and benchmarks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Three observability pillars let coding-agent harnesses evolve autonomously to beat human designs and transfer across benchmarks.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"a2e5b0e9c835855ba0d59c28e90c9105e6239d1a8af3b112cca51ebd8cb8a7c5"},"source":{"id":"2604.25850","kind":"arxiv","version":4},"verdict":{"id":"887645bd-c5cc-42e6-b8be-6c0dde0800bc","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-07T16:12:58.518385Z","strongest_claim":"ten AHE iterations lift pass@1 on Terminal-Bench 2 from 69.7% to 77.0%, surpassing the human-designed harness Codex-CLI (71.9%) and the self-evolving baselines ACE and TF-GRPO. The frozen harness transfers without re-evolution: on SWE-bench-verified it tops aggregate success at 12% fewer tokens than the seed, and on Terminal-Bench 2 it yields +5.1 to +10.1pp cross-family gains across three alternate model families.","one_line_summary":"AHE automates coding-agent harness evolution via component, experience, and decision observability, raising Terminal-Bench 2 pass@1 from 69.7% to 77.0% with transfer gains across models and benchmarks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the three observability pillars sufficiently constrain the action space and provide actionable signal so that the evolution loop produces generalizable improvements rather than benchmark-specific overfitting or noise-driven changes.","pith_extraction_headline":"Three observability pillars let coding-agent harnesses evolve autonomously to beat human designs and transfer across benchmarks."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.25850/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_compliance","ran_at":"2026-05-19T20:43:22.071216Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"ea2b183ffce07ff7ac02fc3848eb521e69fa16beceae61e1f461ea05cac388f6"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}