The Harness, Not the Weights: Where Recursive Self-Improvement Actually Starts

Lilian Weng argues that the near-term path to recursive self-improvement (RSI) won’t run through a model literally rewriting its own weights. Instead, the leverage sits in the ‘harness’—the software layer wrapping a base model that orchestrates how it plans, calls tools, manages context, stores artifacts, and evaluates its own results. Framed as something closer to an operating system than a prompt template, a good harness hides complex orchestration behind a simple interface, and Weng expects its configs, tool interfaces, and protocols to standardize across the industry the way coding agents like Claude Code and Codex already have.

The piece lays out recurring design patterns: goal-oriented workflow loops (plan, execute, test, improve) modeled on Karpathy’s autoresearch repo; using the file system as durable memory so long-horizon runs aren’t bottlenecked by the context window; and spawning inspectable sub-agents and backend jobs whose logs and status persist as files rather than vanishing into transient chat context. A related thread is context engineering—illustrated by Agentic Context Engineering (ACE), which treats context as an evolving playbook of itemized bullets, split across a generator, reflector, and curator, specifically to avoid the ‘context collapse’ that comes from repeatedly rewriting one giant prompt.

The strategic claim is that the harness itself becomes the optimization target, progressing from instruction prompts to structured context, workflows, harness code, and eventually optimizer code as models grow more capable. Mature harnesses enable auto-research loops that improve future models, while smarter models keep harnesses from over-engineering—a mutually reinforcing feedback loop. Weng draws the analogy to prompt engineering: manual tricks faded as instruction tuning improved, but the need to specify goals, constraints, and evaluation never went away. Likewise, many harness behaviors may get internalized into models over time, yet the interface to external tools and context will persist.