Lilian Weng: the harness matters as much as the model

Lilian Weng's new essay makes a simple argument that is easy to miss: when people talk about AI systems improving themselves, they picture a smarter model, but much of the progress is coming from the harness. The harness is the orchestration around the model, everything that turns raw weights into a working system: how it executes steps, calls tools, and stores memory. Weng's claim is that this scaffolding is as important to recursive self-improvement as model capability, and that it is often the part you can improve fastest.

She groups current practice into three design patterns. The first is workflow automation, chaining steps into a repeatable process. The second is persistent file-system memory, giving the agent a place to write things down and read them back across runs. The third is parallel sub-agents that split a task and work at the same time. On top of these she describes ways to optimize the harness itself, including evolutionary search over workflows and self-modifying setups she calls Meta-Harness and Self-Harness, where the system rewrites its own orchestration.

The essay does not sell this as solved. Weng is direct about the open problems: evaluators for research-style tasks are weak, so it is hard to tell whether a change actually helped; memory does not scale cleanly; reward hacking remains a live risk; and human oversight is still needed. That framing, capability and danger in the same picture, is what makes it worth reading rather than another agent-patterns roundup.

Why it matters

If you build agents, this is a clear map of the levers you actually control day to day, since you can change the harness far more easily than the model. It also names the failure modes to watch, so treat the weak-evaluator and reward-hacking warnings as a checklist before you let a system tune its own workflows.