MetaClaw paper: continuous agent learning
MetaClaw published a continuous‑learning approach for deployed LLM agents that extracts skills from failures in real time and fine‑tunes slowly during idle periods — skills alone improved accuracy by 32% on simulated workday benchmarks. The method promises fast closed‑loop fixes without downtime for production agents. (x.com)
MetaClaw’s technical report was posted March 17, 2026 and lists authors including Peng Xia, Jianwen Chen, Xinyu Yang and collaborators from UNC–Chapel Hill, Carnegie Mellon, UC Santa Cruz and UC Berkeley. (alphaxiv.org) The system pairs a fast, inference‑time “skill‑driven” adaptation path that synthesizes reusable behavioral skills from failure trajectories with a slower, gradient‑based opportunistic policy optimizer that performs cloud LoRA fine‑tuning and RL updates. (arxiv.org) Policy updates are scheduled by an Opportunistic Meta‑Learning Scheduler (OMLS) that detects user‑inactive windows via configurable sleep hours, keyboard inactivity and Google Calendar occupancy to trigger training without taking live services offline. (arxiv.org) MetaClaw enforces an explicit versioning and support‑query data partitioning workflow to prevent training‑time data contamination and to preserve a valid meta‑learning signal during continuous operation. (emergentmind.com) Benchmarks run on MetaClaw‑Bench and AutoResearchClaw show the pipeline raised Kimi‑K2.5 task accuracy from 21.4% to 40.6% and improved a composite robustness metric by 18.3% in the paper’s experiments. (huggingface.co) Code and docs were published to a public aiming‑lab GitHub repository with an MIT license and runnable CLI commands (metaclaw setup / metaclaw start); the project emphasizes a proxy‑based architecture that scales to production‑size models without requiring local GPU clusters. (github.com)
