HarnessNet: From Engineered Harnesses to a Unified Neural Controller for Self-Evolving Agents

An LLM agent is shaped not only by its language model, but also by the harness that runs it: the control logic that decides what context the model sees, which proposed commands are executed, when to call the model again, and when to stop. These decisions are central to long-horizon agent performance, yet existing harnesses are typically hand-engineered or obtained through search over harness programs, and remain fixed after deployment. We introduce HarnessNet, a neural feedback controller that learns the harness while keeping the LLM frozen. At each step, HarnessNet reads the interaction history and emits one executable control: request the LLM, execute a previously proposed command, read an environment output, or finish. Controls can reference earlier replies or observations instead of regenerating them, allowing long outputs to be reused exactly. The controller is first trained to imitate an engineered harness, then improved from task success by comparing multiple attempts on the same task, without gradients through the LLM or environment. On AppWorld, HarnessNet matches and then surpasses engineered control. Imitation alone reaches 86.0% success on all 57 development tasks, matching the Meta-Harness baseline under the same budgets. Two reward-driven updates from only 48 training attempts raise success to 91.2%, outperforming both the imitation controller in the same evaluation and the engineered harness. The newly solved tasks mainly come from cases where imitation previously produced invalid controls or exhausted its interaction budget, showing that learning improves how the agent is run rather than the frozen LLM itself.