Treat agents like production infrastructure

The six controls

Every agent I put into production has these six controls before it goes live.

Permissions

Each agent gets its own identity and the narrowest access that does the job. It does not borrow an admin account. You can review and revoke its access like any other service.

Approvals

High-risk actions wait for a person. You decide which actions are high risk: a refund over a limit, a change to a customer record, an email to a client. The approver sees what the agent wants to do and why.

Audit

Every prompt, tool call, decision, and result goes into a record that nobody can quietly edit. When someone asks “why did this happen?”, you have the answer.

Evals

A set of test cases runs before each change goes live. If a new prompt or a new model makes the agent worse, the evals catch it before your customers do.

Observability

You see cost per task, error rate, and cycle time on a dashboard. Alerts go to the people who already run your systems.

Cost control

Each agent has a budget. Limits stop a loop or a bad prompt from running up a large bill overnight.

The method

The method maps to the three services.

1. Assess. Find the workflows where an agent pays off. Score them for fit, risk, and data readiness.
2. Build. Put one agent into production with all six controls. Measure it against the targets we agreed.
3. Operate. Tune it, add workflows, and hand the work to your team.

What I measure

We agree on the measures before the build starts. Usually they are:

• Hours saved on the workflow, each week.
• Error rate compared with the current process.
• Cycle time from request to done.
• Risk that the controls remove or reduce.

If an agent does not move these numbers, we change it or turn it off.

Vendor-neutral

I am independent. I recommend the models, tools, and platforms that fit your workflow and your data rules. I take no referral fees and no commissions.