AI harness

Two teams can use the same model and get very different results. The difference is rarely the model. It is whether one team gave it the right tools, trimmed the context to what matters and added a check before anything reaches a customer. When an AI workflow disappoints, the setup around the model is usually the first place to look.

A language model on its own only turns text into text. To do useful work it needs a setup around it, and that setup is the "harness". It usually has six parts:

• Instructions that say what the job is and what counts as done.
• Tools the model can call, such as a CRM lookup or a send-email action.
• Context, meaning the records and history it is shown for this task.
• Memory of what happened in earlier runs.
• Rules and checks that stop bad actions or catch bad output.
• A place to put the result, such as a record, a queue or a document.

Claude Code is a good example. The model is Claude, and the product around it, with file access, a terminal, permissions and a rules file, is the "harness".

Giving a model every tool at once, so it picks the wrong one. Stuffing in every document it might need, which buries the useful part. Skipping the check because the first ten runs looked fine.

• Start with one job and write down what a good result looks like.
• Add only the tools that job needs.
• Show the model the relevant record and recent history, not the whole system.
• Add a check before any output touches a customer or a system of record.
• Keep a log of every run, so mistakes can be found and fixed.