Review it together, then run it

The agent wrote it and tested it. Before it carries real enquiries, two people read it, and they read different things.

What the business owner reads. The tree view in the editor, where every model step is coloured and every branch is visible: is there a model where a rule should be, is there a rule where a person should be, is the approval where policy says it is? One run in the inspector, step by step, to see what the prospect saw and what the rep saw. A morning's worth of the rep's requests, to see what the job now looks like. None of this needs the JSON.

What the developer checks. In the code view, against a short list. Every agent has limits and a timeout. Every agent has only the tools it needs, and the steps that read untrusted text have none. Every consequential action sits behind a gate. State is typed by output schemas where a model writes it. Every template resolves, which the validator confirms. The connection names are logical, and the deployment binds them.

Then run it. Deploy to production bound to the real CRM. Watch the runs list and the tokens each run used. Ask the analytics in plain words: how many enquiries this week, how many needed a rep, what share was declined and why. Keep the inputs that surprised you and run them again after every change. Publish the workflow to the library so the next team forks it instead of starting from a blank page. Deployments are snapshots, so a bad change is one rollback away. Evaluation suites that score a versioned set of cases against a deployment are still on the way; until then the run records are the evidence.