Paper formalizes AI harness engineering: 11 runtime duties behind coding agents
bibryam · x · 2026-11-15
An arXiv paper by Hailin Zhong and Shengxin Zhu argues autonomous software-engineering capability emerges from a model-harness-environment system, not the model alone. It formalizes the harness as 11 component responsibilities — task specification, context selection, tool access, project memory, task state, observability, failure attribution, verification, permissions, entropy auditing, and intervention logging — operationalized via an H0-H3 ladder and a trace-based evaluation protocol that turns each agent run into an auditable episode package: lower levels yield only a final patch, higher levels add reproduction logs, failure attributions, deterministic checks, and structured verification reports.
More from coding & agent
- Agent frameworks assume Python — what do TypeScript developers actually use? — ConsultantBrannon · 2026-11-15
- Prompt tip goes viral: two rules to make AI-generated replies sound far less robotic — lipeng0820 · 2026-11-15
- Row-Bot drives your desktop with mouse and keyboard — locked down by design — Acceptable-Object390 · 2026-11-15
- Celesto CI undercuts GitHub Actions by 89%: $6.67 vs $60 per 10k minutes — aniketmaurya · 2026-11-15
- Running a customer-emailing agent: inbox ownership vs domain reputation — Not_A_Lender · 2026-11-15
- 16-minute full walkthrough released for the $400/month AI sales outreach workflow — aryanXmahajan · 2026-11-15