1. One path, ugly and real
Build a thin end-to-end slice before you widen. One user, one request, one result, through the real doors: the screen, the agent, the tool, the record. An ugly path that works exposes the integration bug while it is still one bug.
Widening first feels like progress. It is how a proof of concept grows five features and zero traces. Many of those projects stall on missing foundations, not on the choice of model.
2. What has to exist before you add the second path
Five foundations, in an order that compounds. The evaluation cases and the success target come before you tune the model. A trace of each step comes next: the input, the context it retrieved, the tool it called, and the output. Without the trace you cannot debug, and in a regulated setting you often cannot explain the decision either.
Then the data. Agents are less forgiving of bad data than people are. There are two kinds: the content the agent draws on, and the tracking data from the traces. A common silent failure is a document that was updated and never re-indexed. Stale content looks like a stupid model.
3. Keep the orchestration boring
If you need more than one agent, pick the simplest pattern that works. One controller that assigns work is easier to follow than a room of agents reacting to each other. A person at a high-stakes or irreversible step is a pattern, not a lack of ambition. The hardening masterclass is the longer version of how a change earns production.
Treat prompts and model versions like code. Version them, review the change, and keep an audit trail. Check for sensitive personal data before it reaches a model. Put the job on a queue with a retry, a timeout, and a way to ignore a duplicate, so a repeated webhook does not post the same comment twice.
4. When it fails, use the incident you already have
Fold the response into the incident process you already run. Detect on quality and on user complaints, not only on uptime. Diagnose from the trace. Contain by rolling back a prompt or a model version, or by routing to a person. Fix the cause, such as a stale index. Add the failure to the eval set so the next change has to pass it.
A worked shape is a code-review helper. Start with comments a person can accept or dismiss. Cap the volume. A flood of low-value comments trains people to ignore the tool. Measure whether people accept the comment. Optimize for that, not for how many comments the agent can write. Design the architecture yourself first. If Claude designed every choice, you cannot judge the choices.
5. What to try next
Name the one path. Ship it to one person who can open it. Keep a fallback, including sample data, in case a network call fails during a demo. Polish after the path works.
The week-sized version of all ten habits is a starter workflow for this week.