AI Papers Reader

Personalized digests of latest AI research

View on GitHub

Live Supervision Lets AI Agents Recover Mid-Run and Keep What They Learn

Long-horizon AI agents, which perform multi-step tasks in real environments such as shells and code repositories, generate abundant experience as they work. But most self-improvement systems use that experience only after a run ends: they reflect on completed traces, judge final outputs, or update prompts and skills from finished attempts. That post-hoc approach cannot rescue a failing run, and it cannot test newly learned lessons against the very work that produced them. Researchers from the AllSpark team argue that self-improvement should instead be live.

To make that possible, they built PILOT, a supervisor–worker harness. A worker executes the task in an isolated context, while a separate supervisor monitors its trajectory through a live channel. The worker can send notifications, ask questions, or report results; the supervisor can steer the next action or abort the run. When the supervisor sees a reusable procedure or recurring failure mode, it writes that knowledge into a persistent skill library and memory. Later workers load the updated harness. The underlying model is frozen—only the harness evolves.

One trace from Terminal-Bench 2.0 shows how steering works. In the medium-difficulty CoreWars task, a Kimi-K2.6 worker spent more than twenty minutes tuning DAT-clear strategies; its experiments repeatedly tied against multi-process opponents. The supervisor sent a live correction: stop testing synthetic opponents, find a published warrior source, and test it against the actual five opponents. The worker acknowledged the direction, adapted the public Silk Warrior 1.3, and passed every required win-rate threshold, including Stone 93%, Paper 94%, Vampire 84%, Snake 55%, and G2-Clear 53%.

PILOT was evaluated with two frozen backbones, GLM-5.1 and Kimi-K2.6, on three long-horizon benchmarks. In one-shot runs, where each task starts from a fresh harness state, PILOT ranked first in five of six backbone–benchmark combinations. On Terminal-Bench 2.0, its average pass rate was 71.6%, compared with 66.3% for Pi, the strongest single-agent baseline, and 61.8% for Terminus-2, a gap of 9.8 percentage points. On SWE-bench Pro, PILOT averaged 59.9%, 4.4 percentage points above Pi’s 55.5%.

In a 20-iteration self-improvement setting on Terminal-Bench 2.0, every harness began with the same initial skill library and instructions. PILOT’s best observed pass rate rose by 14.6 percentage points with GLM-5.1, from 66.3% to 80.9%, and by 12.4 points with Kimi-K2.6, from 68.5% to 80.9%. Its skill library grew by 21 and 31 skills, respectively. Mean output tokens per evaluated task fell by 42.9% and 47.4%, while successful evaluations per million output tokens rose by 110.3% and 134.0%.

Those gains came with costs. Iterative self-improvement reruns every task many times, so expanding to more benchmarks or backbones would be expensive; the team therefore tested only three benchmarks and two open-weight models. The supervisor and worker also share the same backbone, leaving potential trade-offs of different model pairings unexplored. Manual trace inspection classified only 2.3% of successful GLM-5.1 one-shot runs and 10.6% of successful Kimi-K2.6 runs as aided by live steering, with the share larger on hard tasks than easy ones.

The results do not prove that live self-improvement works in every agent setting. But they point toward a design in which oversight is separated from execution and knowledge accumulates during a run rather than after it—an approach that could make long-horizon agents more recoverable and more efficient without retraining the model.