Dev&mark Engineering Perspectives

Taking an AI Agent from Pilot to Production

A practical guide to tool permissions, evaluation, approvals and operational ownership for business agents.

The Short Answer

Moving an agent into production requires more than a working demo: restricted tools, representative evaluations, explicit approvals, monitoring, recovery and an accountable operator all need an agreed design.

Original Perspective

A Practical Guide for Technology Leaders.

Dev&mark Editorial · Published November 20, 2026

Define the Action Boundary

List what the agent may read, prepare and change. Separate an information lookup from an irreversible action. Each tool needs a clear input contract, an expected result and an error path. The model should not decide its own permission level. Start with fewer actions and expand only when the workflow supports it.

Design Approval as a Real State

An approval should show the proposed action, relevant context and the effect of proceeding. The system must know whether approval was granted and whether the approved proposal is still current. If the information changes after review, the agent may need a new decision. A message that says approval is needed is not an enforced control.

Test Failures and Ambiguity

Evaluate missing information, conflicting instructions, unavailable systems and malformed tool results. Include attempts to access information outside the user’s permissions. Define stop and escalation conditions before running unattended tasks. Test whether a retry could repeat a payment, message or record change; not every failed response means the action did not happen.

Make Behavior Inspectable

Agree trace information that explains the task, selected tools, outcomes and review states. Keep logging proportionate to the need and restrict sensitive content. Product and operations owners should be able to investigate a failure without relying on the agent’s own summary of what it did. Decide who owns that investigation.

Treat Changes as Releases

A different model, prompt, retrieval source or tool contract can change behavior. Run representative evaluations before rollout, document the change and maintain a recovery path. Review operating cost and service limits for the expected workload. Production readiness is a set of responsibilities and evidence, not a model name.

Use a Readiness Decision

Before launch, confirm the owner, allowed actions, approval behavior, evaluation evidence, failure response and support scope. If one of these is unresolved, a supervised pilot may remain the appropriate operating mode. It is better to narrow a workflow than to describe unsupported autonomy as a finished capability.

Put It into Practice

Connect the Guide to Your Next Step.

Discuss your workflow, engineering context and the evidence needed to make a useful decision. The service scope and responsibilities are agreed around your situation.

Explore the Related Service
Continue Exploring

Connect Strategy, Engineering and Delivery.

US Presence · International Delivery

Build the Next Step with Dev&mark.

Tell us about your goals, existing systems and delivery needs. Start with a 30-minute technical conversation.

Book a Technical Call