The authors introduce Growing Harness, a training paradigm that learns an agent's control structure from task feedback rather than relying on standard context-heavy harnesses. By converting recurring control decisions into executable code and reserving LLM calls for semantic reasoning, the method aims to create reusable specialist agents.
- Growing Harness uses a strategy-free scaffold with fixed interfaces, localizing failures via function-level execution traces.
- An optimizer repairs windows of failures jointly, while a success-first gate rolls back edits that harm prior capability.
- Across BrowseComp-Plus and WebArena-Verified with models from 4B to 120B parameters, it achieved the highest mean success in five of six settings.
- It reduces LLM calls by 76.0-91.8% and deployed-agent inference cost by 74.4-98.6% relative to a Tool-Calling agent.
- On WebArena-Verified, success remains stable at 44.7-45.3% across model scales, whereas Tool-Calling drops to 6.7% with the 4B model.
Persistent program growth moves recurring control out of model context into low-cost code, allowing agents to remain effective even when deployed with smaller models.