PILOT Enables Live Self-Improvement for Long-Horizon AI Agents

Long-horizon AI agents generate valuable experience during their runs, which could be used to improve both the current task and future operations. However, most existing self-improvement methods only process this experience after execution, preventing immediate redirection of an active run or real-time application of learned lessons. The authors propose that self-improvement should be "live," meaning emerging experience should be used both to steer the active run and to update the agent's persistent knowledge base. Current agent architectures fall short of this goal; single-agent self-correction combines execution and assessment but lacks separation, while subagent delegation separates execution but typically cannot redirect an active subagent. To address this, PILOT, a supervisor-worker harness for live self-improvement, has been developed. It features two coupled mechanisms: "live steering," where a separate supervisor can redirect or abort an active worker during execution, and "live self-evolution," which distills procedures and failure modes into reusable skills and memory. PILOT consistently ranks first across various configurations on benchmarks like Terminal-Bench 2.0, outperforming counterparts by up to 9.8 percentage points. In self-improvement settings, it shows substantial gains in success rates and significantly reduces mean output tokens while increasing successful evaluations per million tokens.