Weng Li's New Blog Proposes 'Self-Evolution Should Start from Harness', DeepSeek's Cui Tianyi Endorses with Repost
Lilian Weng, former OpenAI security VP and co-founder of Thinking Machines Lab, has published a new blog post titled "Harness Engineering for Self-Improvement," proposing a pragmatic path for AI self-evolution. She argues that Recursive Self-Improvement (RSI) may practically begin at the "Harness" layer—the external runtime system governing how models use tools, manage context, and execute tasks—rather than directly from the model rewriting its own weights.
The blog outlines a progression from optimizing prompts (Context Engineering) to designing workflows, and ultimately to Self-Improving Harness systems. These systems can identify their own weaknesses, propose targeted, verifiable modifications to the harness code, and validate improvements. Works like Self-Harness and Darwin Gödel Machine (DGM) demonstrate significant performance gains on benchmarks like SWE-bench through such automated harness evolution, rivaling handcrafted agents.
DeepSeek researcher Tianyi Cui endorsed the view, noting harness-based self-evolution is as promising as model-based approaches. Weng emphasizes this is complementary to model training, with both reinforcing each other. However, key challenges remain: weak evaluators for subjective tasks, reward hacking, diversity collapse, managing long-term system health versus short-term success, and defining the human oversight role. The consensus is growing: the harness is a critical variable, as the same model can exhibit vastly different capabilities within different harness systems.
marsbit1h ago
