Build around real work. Test before wider use.

Understand the job

Map the current work with its owner. Review representative inputs, existing software, access permissions, and the cost of getting it wrong.

Workflow map + system inventory

Agree the outcome, exclusions, owner, and native-versus-custom recommendation.

Design the boundaries

Specify tools, source access, actions, approval gates, and the path for incomplete or hostile inputs. Decide how to test behavior before building broadly.

Permission matrix + evaluation cases

Approve the behavior, stopping rules, and acceptance tests.

Build and connect

Implement the agent and its interface. Validate tool arguments, handle duplicate triggers, and test failed calls and partial completion in an isolated environment.

Integrated system + test evidence

Demonstrate the agreed cases, including actions that must not happen.

Release with an owner

Begin in a sandbox or read-only mode, then reviewed actions. Broader permissions require a separate decision backed by results.

Release checklist + operating runbook

Confirm ownership of alerts, pause controls, exceptions, and rollback where possible.

Improve with evidence

Review accepted outputs, errors, costs, and team feedback. Re-run evaluations when prompts, models, sources, or connectors change.

Regression results + change record

Expand scope only when the evidence and responsible owner support it.