PILOT Brings Live Steering and Self-Evolution to Long-Horizon Agents

Introduction

Long-horizon agents often fail because an early mistake compounds over many later steps. A conventional self-improvement loop waits until execution has finished, reviews the trajectory, and applies the resulting lesson to a future run. That can improve subsequent attempts, but it cannot rescue the run that produced the evidence in the first place.

PILOT argues that improvement should happen while the task is still unfolding. Its supervisor–worker harness separates task execution from oversight. The worker carries out the task, while an independent supervisor monitors the trajectory and can intervene before the run ends. At the same time, procedures and failure modes discovered during execution are distilled into reusable skills and persistent memory.

Core mechanisms

• Live steering: The supervisor can redirect an active worker when its behavior appears to be moving away from the objective. It can also abort a run rather than allowing an unproductive trajectory to continue.
• Live self-evolution: Runtime experience is converted into reusable procedures, skills, and memory instead of being treated only as a final postmortem.
• Decoupled oversight: Unlike single-agent self-correction, execution and trajectory assessment are placed in separate roles. Unlike ordinary subagent delegation, the supervisor is designed to influence a worker that is already running.

Reported results

The paper evaluates PILOT with two frozen backbones, GLM-5.1 and Kimi-K2.6, across three benchmarks and six total configurations. PILOT ranks first in five of those configurations. On Terminal-Bench 2.0, it improves over counterpart harnesses by as much as 9.8 percentage points.

In the self-improvement setting, the reported gain is 14.6 points with GLM-5.1 and 12.4 points with Kimi-K2.6. Mean output tokens fall by 42.9% and 47.4%, respectively. Successful evaluations per million output tokens rise by 110.3% and 134.0%. Taken together, these figures suggest that intervention can both improve outcomes and reduce prolonged exploration along unproductive paths.

Why it matters

PILOT reframes an agent harness as an active control loop rather than a passive wrapper around a language model. The supervisor is not merely a grader operating after completion; it can alter the current trajectory and help improve the system that will handle later tasks. This is particularly relevant to terminal-based work and other workflows where mistakes become expensive only after many dependent actions.

The approach also raises practical questions. A supervisor must distinguish a genuine strategic failure from a temporary setback, and an intervention can itself make a viable trajectory worse. The supplied material does not specify detailed rollback behavior, intervention latency, or the cost of introspection. Those measurements will be important for judging whether live self-improvement remains stable outside benchmark settings. The authors state that code will be released on GitHub, which may make these design choices easier to examine.

Source: Hugging Face Daily Papers