Grow the harness, not the context: recurring control becomes code
Sep 22 Li et al.: failure-guided Growing Harness turns a strategy-free scaffold into reusable executable control — 76–92% fewer LLM calls, held-out gate rollback, success holds across 4B–120B on WebArena.
Source: Grow the Harness, Not the Context: From Strategy-Free Scaffolds to Reusable Specialist Agents
Li, Li, Zhao, Ye, Li, Xu, and Gao ask a factory question most overnight loops dodge: if an agent keeps seeing the same family of tasks, why does every run reconstruct the same control decisions inside a growing model context? Growing Harness is a failure-guided training paradigm that learns the harness itself from a strategy-free scaffold — fixed model and tool interfaces, no hand-designed ReAct-style controller. Function-level execution traces localize each failure to a bounded code surface; an optimizer jointly repairs a window of failures; a success-first held-out gate rolls back repair sequences that harm prior capability. Accepted edits accumulate in one shared program so control structure emerges from task feedback.
The receipts matter for continuous factories. Across BrowseComp-Plus and WebArena-Verified with deployment models from 4B to 120B, Growing Harness hits the highest mean success in five of six settings (within 0.7 pp on the sixth). Versus a Tool-Calling agent it cuts LLM calls by 76.0–91.8% and online inference cost by 74.4–98.6%. On WebArena-Verified, success stays 44.7–45.3% across model scales while Tool-Calling collapses to 6.7% at 4B — reusable code supplying the browser control small models otherwise re-derive every turn. Ablations show trace-local edits, multi-failure windows, and gate rollback each lift final success; without the gate, validation success rose then fell as brittle repairs landed.
Read next to empirical harness-component studies and other self-editing control planes: the shared claim is that the durable artifact is the control program, not a longer chat. Caveats the authors own: value depends on reusing the learned harness enough to pay offline optimization, and generated code still needs sandboxing and permission boundaries.
Why it matters for code review
Self-improving factories are not “let the agent rewrite prompts overnight” — they are merge-gated program growth where recurring verify/recover/stop logic leaves the context and becomes reviewable code. Treat harness diffs like product diffs: held-out success cannot regress, edits stay trace-scoped, and morning inventory is the accumulated controller plus evidence it still passes the gate. That is how modest verified momentum compounds past peak-session heroics without burning the night re-deriving the same control loop in tokens.