A Decent Model With a Great Harness Beats a Great Model With a Bad Harness

The most useful thing I've read about agent engineering this month isn't about models. It's about harnesses.

Addy Osmani published Agent Harness Engineering, and it's the kind of post where every paragraph contains something I want to steal for my own projects. The core argument: the harness you wrap around an AI agent matters more than which model powers it. I've been feeling this in my own work with Claude Code in my personal projects, but Osmani gives it a name and a set of patterns.

The Ratchet Pattern is my favorite. Every time an agent makes a mistake, you convert that specific failure into a prevention mechanism. A rule in AGENTS.md, a pre-commit hook, a verifier subagent. The harness gets stricter over time but only in response to real failures, not hypothetical ones. Your AGENTS.md should be under 60 lines, and every rule should trace back to a specific past failure. If you can't point to the incident that created the rule, delete the rule.

Ralph Loops solve a problem I've hit repeatedly: agents that drift off task in long-horizon work. When the agent exits prematurely or loses context, you re-inject the original prompt into a fresh context window. Simple idea. Hard to get right without a framework for detecting premature exit.

Sprint Contracts flip the usual agent interaction model. Instead of giving the agent instructions and hoping for the best, the agent negotiates completion criteria before writing any code. "I'll consider this done when tests X, Y, Z pass and the API responds with 200 on this endpoint." You agree to the contract, then the agent works. If Shopify's River (story four) is any indication, this kind of structured constraint is exactly what makes agent-written PRs trustworthy at scale.

Osmani also published a companion piece on Agentic Engine Optimization, covering the other side: structuring your docs so AI agents can actually consume them. Create llms.txt at your domain root as an agent sitemap (under 5K tokens), track token counts as a first-class metric, write skill.md files declaring service capabilities. This is real work that pays off immediately if agents are using your APIs.

The operational insight that stuck with me: use hooks at lifecycle points. Silent on success. Errors surface for agent correction. Don't log everything. Only log what the agent needs to fix.