Harness Handbook maps agent behaviors back to source code and lifts plan accuracy
omarsar0 · x · 2026-07-23
The paper Harness Handbook: Making Evolving Agent Harnesses Readable, Navigable, and Editable proposes a behavior-centric way to localize code behind agent behaviors.
What it does
- Builds a three-level map from runtime behaviors to source locations using static analysis plus LLM-assisted structuring.
- Introduces BGPD (Behavior-Guided Progressive Disclosure), which guides coding agents from a system overview down to relevant stages, functions, and files, then checks candidates against the current source.
Reported results
On 60 modification requests across Codex and Terminus-2:
- Planning win rate rose from 28.3% to 38.3% on one benchmark and 26.7% to 45.6% on the other.
- Planner token usage dropped 12.7% and 8.6%.
- File- and symbol-level F1 improved in all 24 comparisons against GPT-5.5 and Opus 4.8 reference plans.
- Complete localization misses fell by up to 25.9 points.
The authors argue that evolving large agent harnesses is not just about generating edits, but also about finding where those edits should happen.
More from coding & agent
- LangChain maps the governance checklist teams need before agents hit production — LangChain · 2026-07-23
- Agent memory works better as a statement graph with provenance and append-only history — chrislally · 2026-07-23
- Pond pitches outcome-based agents that only charge when the work is delivered — ThePeterMick · 2026-07-23
- Factory ships model routing after costs drop 25% on task-by-task model selection — matanSF · 2026-07-23
- Claude Code has a hidden /radio command for surprisingly decent background music — ryanbed · 2026-07-23
- Claude Code, Cursor, CodeRabbit and MCP form a one-line AI coding stack — IndraVahan · 2026-07-23