On this page
Learning objectives
- Define deployment, secrets, access, and change controls
- Design recovery for partial failure and external side effects
- Measure full operating cost per useful outcome
Prerequisites
A passing evaluation set · A named service owner · Staging environment
Answer first
Canonical definition
A production agent is a versioned software service with an owned mandate, constrained identity and tools, observable state, tested failure behavior, release controls, and an operating cost tied to useful outcomes.
Operating context
Why it matters
The production boundary includes permissions, data retention, vendor behavior, queues, humans, and downstream systems—not only the model endpoint.
Cost must include review, retries, failures, maintenance, and infrastructure before an automation claim is credible.
Method
Step by step
- 01
Harden identity and access
Use separate service identities, least privilege, short-lived secrets, environment boundaries, and access reviews.
- 02
Control releases
Version prompts, tools, policies, schemas, and models; require checks and approval before production changes.
- 03
Design recovery
Make side effects idempotent, persist checkpoints, support pause and replay, and document manual continuation.
- 04
Measure unit economics
Calculate cost per successful useful task including infrastructure, model, tools, review, rework, and on-call.
Hands-on lab
Write a production readiness review
Assess one agent against deployment, security, privacy, reliability, evaluation, support, and cost gates.
Deliverable
A go/no-go checklist, risk register, runbook, rollback plan, dashboard, and named owners.
What breaks
Common failure modes
- F1Sharing a human employee credential with an agent
- F2Deploying prompt changes without regression tests or rollback
- F3Counting only token charges when comparing with human work
Beyond the demo
Production notes
- Rehearse credential revocation, vendor outage, poisoned input, repeated side effect, and manual failover.
- Review model and platform data-retention terms against the actual data classes used by the workflow.
Further reading
Sources
Related Tenten resources