A mediocre agent inside a strong harness outperforms a stronger agent inside a messy one

The central lesson of Karpathy’s Autoresearch is that the harness is the product, not the agent. The agent edits one file, chases one metric, operates within one fixed harness, and advances only when the score improves — and that’s not a limitation but the reason the system can run for hours without dissolving into noise. As the analysis puts it: “A mediocre agent inside a strong harness can outperform a stronger agent inside a messy one.”

This reframes the AI capability conversation. A lot of builders focus on model intelligence in isolation, but Autoresearch shows that the surrounding machinery matters just as much: how work is launched, how failures are handled, how progress is measured, how bad paths are rolled back. Intelligence location — code vs prompts — determines system fragility and flexibility generalizes this: the best systems put deterministic constraints in code and reserve prompts for judgment calls — the harness IS the code-driven intelligence layer. This is Verification is the single highest-leverage practice for agent-assisted coding taken to its logical conclusion — verification isn’t just a quality check, it’s what makes the entire autonomous loop viable. The harness compounds over time through Compound engineering makes each unit of work improve all future work, and the constraints themselves become capabilities in the spirit of Declarative beats imperative when working with agents. The practical proof is in Autonomous coding loops need small stories and fast feedback to work — the Ralph pattern works because the harness is tight, not because the agent is smart. AutoAgent’s Meta-agents that autonomously optimize task agents beat hand-engineered harnesses on production benchmarks provides the strongest evidence yet: a meta-agent autonomously iterating on a task agent’s harness hit #1 on two production benchmarks, beating every hand-engineered entry.