Lilian Weng's Deep Dive: The Key to Agent Self-Improvement Lies in Harness Engineering
aigclink · x · 2026-08-09
Former OpenAI research lead Lilian Weng explores practical paths to Recursive Self-Improvement (RSI) in a deep blog post. She argues that short-term RSI relies not on directly rewriting model weights, but on optimizing the Harness—the deployment system connecting the model to the real world. This layer is where top coding agents like Claude Code and Codex build their competitive moats.
Weng observes that the toolset for coding agents is converging (e.g., File System, Shell, MCP/Skills). The future battleground is Harness optimization, categorized into three main directions:
- Context Engineering: Treating context as an evolving playbook rather than an ever-growing prompt, and separating context management mechanisms from the actual content.
- Workflow Design: Optimizing execution paths using graph search (like MCTS) and strict evidence traceability.
- Self-Improving Harness: Enabling agents to directly modify their own underlying codebase. For instance, the Darwin Gödel Machine project leveraged this to boost its SWE-bench Verified pass rate from 20% to 50%.
Related event: Lilian Weng Explores Harness Engineering for AI Self-Improvement(3 posts)→
More from coding & agent
- DIRECT: retrieval-based next-question suggestions for conversational agents — burkov · 2026-10-16
- Claude Code and Codex now autocomplete the user's next action — tokenbender · 2026-10-16
- Hermes Agent ships official Reverse Engineer Anything plugin with one-command setup — Teknium · 2026-10-16
- MLExpert: 35B MoE Agent Trained to Autonomously Handle Machine Learning Engineering — Shangheng Du · 2026-10-16
- Cowork-Gym: Alibaba Synthesizes 64.9K Agent Work Environments, Boosting Qwen by 8.7 Points — Anni Zou · 2026-10-16
- MMSearch-OPD: Joint RL and On-Policy Distillation for Long-Horizon Multimodal Search Agents — PekingUniversity · 2026-10-16