The model is the commodity now
Models converge. Harnesses don’t. In a Databricks harness benchmark from August 2026, the same model at the same thinking effort produced more than 2× cost differences across harnesses while quality stayed flat — and the simple, minimal harnesses usually won.
That flipped how I think about agent tooling. The value isn’t in picking the biggest model, it’s in everything wrapped around it: the system prompt, the tool surface, the context budget, and the agent loop that ties them together. That’s harness engineering.
Keep the surface area small
A bloated system prompt is a tax you pay on every single turn. It burns prompt cache, it dilutes attention, and it makes the agent dumber in the exact areas you care about.
- Treat every built-in tool as a feature you have to justify. Four tools —
read,write,edit,bash— cover most coding work. - Keep the system prompt under ~1,000 tokens. If a capability matters, add it as a skill loaded on demand, not as prose in the base prompt.
- Measure the context budget. Bigger isn’t better; a leaner window with the right files in it beats a stuffed one.
// Skill: loaded only when needed, keeps the base prompt minimal
export const skill = {
id: "migration-audit",
load: () => fs.readFile("skills/migration-audit.md", "utf8"),
};The loop is the product
The real work happens in the agent loop: how tools are exposed, how errors feed back, how history gets compacted. Engineering that loop — retries, checkpointing, cheap models for extraction and escalation to a big one only on failure — is where the leverage is.
Harness engineering is boring on purpose. It’s fewer tools, smaller prompts, and a tighter loop. The model changes every quarter; your harness is the thing you actually ship.