Harness engineering is the next agent advantage

Lilian Weng argues that near-term recursive self-improvement may come from improving the harness around a model: the runtime that manages planning, tools, context, files, subagents, evaluation, and iteration. Why it matters: This reframes agent progress as systems design, not just model choice. The winning workflows will likely have durable state, explicit process management, inspectable subagent outputs, verification gates, and harnesses that can improve from their own trajectories.

Try this: Audit one agent workflow as a harness: list its trigger, loop, context store, file artifacts, subagents/jobs, permission boundaries, evaluation checks, and recovery path after interruption. Upgrade the weakest link before adding more model calls.