The model is not the harness
It is tempting to think capability lives in the model. Most of the time, the difference between an agent that flails and an agent that ships is the harness: what tools it can call, what it can see, what it remembers, and how its work gets checked.
Two teams pointing the same model at the same task get wildly different results, and the delta is almost always harness design.
What a good harness gives the model
A well-designed Claude Code harness supplies the things a raw model lacks:
- Tools — the ability to read files, run commands, search, and observe real results instead of guessing.
- Context — the right slice of the codebase in the window at the right time, and not the whole thing.
- Memory — durable state across runs, so the agent isn’t starting from zero every session.
- Feedback loops — tests, typechecks and reviews that let the agent see when it’s wrong and correct itself.
- Guardrails — permissions and boundaries so a capable agent stays a safe one.
Why it’s worth getting right
Harness design compounds. Every improvement to how the agent sees, remembers and verifies its work pays off on every task afterwards. It is the highest-leverage engineering in applied AI, and it is still mostly unwritten.