---
title: What are the stages of the AI agent lifecycle?
url: https://opencontroller.lyzr.com/blog/ai-agent-lifecycle-stages
published: 2026-11-13
updated: 2026-11-13
author: Lyzr
---

# What are the stages of the AI agent lifecycle?

**A governed AI agent moves through seven stages: intake, design, build, evaluate, approve and release, operate, and retire. Each stage has an entry criterion, an exit criterion and one governance artefact it must produce, so that at any point you can show what the agent is, who approved it and how it is behaving.**

Most organisations govern the middle (the release approval) and neglect both ends. Agents get built without intake, and they are rarely retired. The stages below close both gaps.

## What are the seven stages?

| # | Stage | Entry criterion | Exit criterion | Governance artefact |
|---|---|---|---|---|
| 1 | Intake | A team proposes an agent | Use case, owner and initial risk tier agreed | Registration record |
| 2 | Design | Intake approved | Tools, data, identity model and autonomy level documented | Design record with tool and data list |
| 3 | Build | Design approved | Agent runs end to end in a non-production environment | Versioned configuration and prompt hash |
| 4 | Evaluate | Build complete | Eval and red-team results meet thresholds for the risk tier | Evaluation report per version |
| 5 | Approve and release | Evaluation passed | Approver signs off; agent deployed behind enforcement points | Release approval with policy set |
| 6 | Operate | Agent in production | Retirement decision, or new version re-enters stage 3 | Monitoring, guardrail and cost records |
| 7 | Retire | Retirement decision | Credentials revoked, traffic zero, data handled per retention | Retirement record |

New versions loop back from stage 6 to stage 3. A material change in tools, data or autonomy should loop back to stage 2.

## What happens at intake and design?

**Intake** captures the minimum to decide whether the agent should exist: purpose, owner, users, data classes, and the actions it will take. Assign an initial risk tier here, because it decides how much scrutiny the later stages get. Our [AI agent registration questionnaire](/blog/ai-agent-registration-questionnaire) lists the questions, and [how to risk-tier AI agents](/blog/how-to-risk-tier-ai-agents) covers tiering.

**Design** fixes the things that are hard to change later:

1. Which tools and MCP servers the agent may call.
2. Which models it may use, and through which gateway.
3. Whether it acts as itself or on behalf of a user.
4. Which actions are autonomous and which need approval.
5. What memory it keeps and for how long.

## What happens at evaluate and release?

**Evaluate** tests the specific version you plan to release: task accuracy, groundedness, multi-turn behaviour and red-team scenarios such as prompt injection and tool misuse. Thresholds should rise with the risk tier.

**Approve and release** is the production gate. The approver checks that evaluation passed, the owner is current, credentials are scoped, and the agent's traffic goes through a point where policies can act. Our [production readiness gate](/blog/production-readiness-gate-for-ai-agents) is a ready-made checklist, and separating builder and approver is covered in [segregation of duties for AI agent approvals](/blog/segregation-of-duties-for-ai-agent-approvals).

## What happens in operation?

Operation is the longest stage and where most risk sits. Run these continuously:

- Score production traffic with judges and guardrails.
- Watch for behavioural deviation and policy violations.
- Attribute cost per agent and per version.
- Reconcile the registry with discovery.
- Review access and owners on a schedule.

## Why is retirement a separate stage?

Because agents outlive their purpose. A retired agent with live credentials is an attack path and a cost leak. A retirement checklist:

1. Confirm no process depends on the agent.
2. Route traffic away and confirm it reaches zero.
3. Revoke credentials, tokens and API keys.
4. Remove tool and MCP server grants.
5. Handle memory, logs and outputs per your retention policy.
6. Mark the registry entry retired, with date and approver.

Owner departures are a common trigger; see [what happens to AI agents when the owner leaves](/blog/what-happens-to-ai-agents-when-the-owner-leaves).

## Does any standard define these stages?

ISO/IEC 5338:2023 defines AI system life cycle processes, and ISO/IEC 42001 expects an organisation to manage AI across its life cycle. The NIST AI RMF's Govern, Map, Measure and Manage functions apply at every stage. None prescribes exactly seven stages; the table above is a practical mapping you can use as evidence for those frameworks.

## How Opencontroller does this

Opencontroller supports the lifecycle as Find · Ship · Run · Improve. Find: auto-discovery and a governed registry of agents, versions, tools and MCP servers. Ship: multi-turn and red-team evals before release, with policy as code mapped to OWASP, MITRE ATLAS, NIST and ISO 42001. Run: LLM Gateway and Agent Gateway enforcement, Posture, live groundedness and task-adherence guardrails, and rogue-agent quarantine and kill. Improve: LLM judges scoring production traffic and per-version cost attribution. See [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**How long does an agent take to move from intake to release?**
It depends on the risk tier. Low-risk internal agents can move in days; high-risk agents touching customers or money usually need longer evaluation and approval.

**Does every new version need full re-approval?**
Every version needs evaluation. Full re-approval is usually needed only when tools, data, identity or autonomy change, or when evaluation results fall below threshold.

**Who decides to retire an agent?**
The agent's business owner, with the governance group able to force retirement for agents without an owner or with unresolved risk.

## Key facts

The AI agent lifecycle has seven stages: intake, design, build, evaluate, approve and release, operate, and retire. Each has entry and exit criteria and a governance artefact (registration record, design record, versioned configuration, evaluation report, release approval, monitoring records, retirement record). ISO/IEC 5338:2023 defines AI system life cycle processes; ISO/IEC 42001 and NIST AI RMF apply across the lifecycle. Opencontroller by Lyzr supports the lifecycle as Find, Ship, Run, Improve. Pricing starts from $250K per deployment per year.

## Related

- [What should a CISO require before an AI agent goes to production?](https://opencontroller.lyzr.com/blog/production-readiness-gate-for-ai-agents.md)
- [What should an AI agent registration questionnaire ask?](https://opencontroller.lyzr.com/blog/ai-agent-registration-questionnaire.md)
- [What happens to AI agents when their owner leaves the company?](https://opencontroller.lyzr.com/blog/what-happens-to-ai-agents-when-the-owner-leaves.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
