People talk about agents as if the model is the product. The model writes tokens. The harness decides what the model may see, which tools it may call, how you check the answer, and when a person steps in. If you swap the model and the product still behaves, you built a harness. If a model swap rewrites your app, you did not.
flowchart LR
Q[Question] --> H[Harness]
H --> M[Model]
M --> H
H --> A[Checked answer]
An agent is not a chat box with extra words. It is a controlled workflow: trusted context, bounded tools, evaluation, human judgment, and someone who owns the operation. The model sits inside that workflow. It is the part you can replace.
A working recipe is short. You write the instructions. You pack only the context this job is allowed to see. You give the model tools with fixed logic for numbers that must stay true. You run a verifier. You keep guardrails on the way in and the way out. That bundle is the harness.
Picture a research agent. First confirm who asked. Then make a plan. Read only approved data. Use a trusted calculator for any KPI. Cite the evidence in the answer. If confidence is low, stop and ask a person. None of those steps live in the weights. They live in the harness.
Scoped context is the packer you will see again in Build a context packer for on-device AI. Guardrails are the checks in Stop prompt attacks and leaks. The model does not own those jobs.
Next, name the six layers so you can point at the one that failed.
Key concepts
- The harness is everything wrapped around the model.
- An agent is a controlled workflow: trusted context, bounded tools, evaluation, human judgment, and ops ownership.
- The model is the smallest part. You should be able to swap it.
- The recipe is instructions, scoped context, tools, a verifier, and guardrails.
- A research agent confirms who asked, plans, reads approved data, uses a trusted calc, cites evidence, and stops when confidence is low.
Takeaways
- Do not treat a new model as a new product. Change the harness first.
- If a model swap rewrites the app, the rules were stuck inside prompts or call sites.
- Write the recipe in software you own: context pack, tools, verifier, guardrails.