The agentic development lifecycle, in seven stages.

Plan → Architect → Prompt → Build → Review → Deploy → Iterate. Seven stages that keep agentic projects fast without sacrificing governance. Every stage has a specific output and an explicit handoff.

Visual lifecycle diagram

Think of this as a loop, not a straight line. Every deployment creates feedback that should restart planning with better context.

After Iterate, feed learnings back into Plan.

Stage-by-stage outputs

Each stage resolves one question. Skip the question, skip the stage, and you end up re-doing work two stages later.

01. Plan

Objective. Define the business outcome and measurable success criteria.

Required deliverable. Scope brief with one target metric and one failure threshold.

02. Architect

Objective. Map agents, tools, permissions, and data boundaries.

Required deliverable. System diagram with ownership and guardrail contracts.

03. Prompt

Objective. Design reusable prompt specs per role and edge-case behaviour.

Required deliverable. Versioned prompt pack with evaluation fixtures.

04. Build

Objective. Implement workflows, memory adapters, and runtime controls.

Required deliverable. Working pipeline with instrumentation hooks.

05. Review

Objective. Run QA, policy checks, and human sign-off checkpoints.

Required deliverable. Approval log plus rollback playbook.

06. Deploy

Objective. Release safely with monitoring and alert thresholds enabled.

Required deliverable. Production release with SLO and incident triggers.

07. Iterate

Objective. Improve prompts, flows, and economics from live feedback.

Required deliverable. Weekly optimisation backlog tied to business impact.

Frequently asked

Where do most teams break?

Between Build and Review, when teams skip explicit validation criteria. Review gates are what protect production velocity.

How long should one full lifecycle run take?

A focused pilot can complete one full loop in one to three weeks, depending on integration complexity and compliance needs.

Can this work for solo operators?

Yes. The stages still apply, but one person may hold multiple roles. The important part is preserving stage boundaries and measurement rigour.

How do teams keep this measurable over time?

Track one primary business KPI, one quality KPI, and one cost KPI for each loop, then review trend drift during every Iterate stage.