Grow the Harness, Not the Context: From Strategy-Free Scaffolds to Reusable Specialist Agents
A learned harness moved recurring control into reusable code, cutting model calls and inference cost while keeping web-agent performance stable across 4B–120B deployment models.
Growing Harness starts without a task-solving controller, then uses execution traces, joint repair, and rollback gates to accumulate reusable control code. Across two web-agent benchmarks and models from **4B to 120B**, it led **five of six** settings.
For recurring agent workflows, treat repeated planning and recovery logic as code that can be learned, tested, and retained. The reported harness reduced LLM calls by **76.0–91.8%** and inference cost by **74.4–98.6%** versus a tool-calling agent.
Growing Harness starts without a task-solving controller, then uses execution traces, joint repair, and rollback gates to accumulate reusable control code. Across two web-agent benchmarks and models from **4B to 120B**, it led **five of six** settings. For recurring agent workflows, treat repeated planning and recovery logic as code that can be learned, tested, and retained. The reported harness reduced LLM calls by **76.0–91.8%** and inference cost by **74.4–98.6%** versus a tool-calling agent. The evidence comes from BrowseComp-Plus and WebArena-Verified, not general software work. The harness also requires trace-level localization, repair optimization, and held-out rollback checks; the abstract does not quantify their training cost.
This shifts harness engineering from manually specifying controls for each workflow to learning reusable control code from execution traces, with rollback checks limiting regressions. It gives benchmark evidence that retained scaffolding can reduce calls and cost, while narrowing the claim to two web-agent settings and leaving harness-training cost unquantified.