ADLC · Slide 3 · Six Stages

ADLC Six Stages

Intent to Deployment. The Agentic Development Lifecycle defines six controlled stages where agents execute and humans govern progress through evidence, review, and release authorization.

Explore the six stagesNext: Five Gates
ADLC six stages slide
Intent · Specification · Planning · Execution · Verification · Operation

The six-stage journey

Define → Specify → Plan → Build → Verify → Operate

ADLC is not a waterfall process. It is a controlled trust-building framework where agents earn autonomy through verification at every stage.

A

Intent & Framing

Define the goal, scope, business need, and success criteria.
B

Specification

Translate intent into testable acceptance criteria and constraints.
C

Planning

Decompose the work into a task graph, dependencies, and guardrails.
D

Execution

Agent builds, tests, and self-corrects within approved boundaries.
E

Verification

Create evidence, run tests, review conformance, and prove readiness.
F

Operation

Deploy, monitor, learn from production, and feed back into the cycle.

Key takeaway

ADLC limits the blast radius of agent errors by making every stage prove readiness before the next one begins.

Agents move fast, but progress remains governed. Humans approve framing, specifications, guardrails, evidence, and release decisions.

DefineDesignPlanBuildVerifyOperateLearnRepeat

Stage details

Agent and human responsibilities by stage

Each stage defines outputs, agent action, human authority, and the evidence needed to move forward.

A

Intent & Framing

Define the goal before building anything

The first stage establishes what is being built and why. The agent quickly generates candidate framings, but the human approves the final definition so the work aligns with real business and user needs.

Goal statementSuccess criteriaScope boundaryStakeholder sign-off
AgentDrafts problem framings, suggests success criteria, and surfaces ambiguity for review.
HumanApproves final problem framing and success criteria. This decision remains human-owned.
B

Specification

Testable acceptance criteria with constraints

The approved intent becomes machine-verifiable requirements. Specifications must be testable because vague requirements cannot be evaluated by agents or gates.

Acceptance criteriaConstraintsEdge casesNon-functional requirements
AgentGenerates measurable criteria, constraint lists, and edge-case inventories.
HumanReviews completeness, clarifies ambiguity, and accepts the final specification.
C

Planning

Task graph, dependencies, and guardrails

Planning turns specifications into execution paths. The agent decomposes work; humans define guardrails, autonomy levels, and boundaries for safe execution.

Task graphDependency mapGuardrailsAutonomy level per task
AgentCreates task sequence, dependency graph, effort estimates, and autonomy recommendations.
HumanSets boundaries, confirms the plan, and defines what the agent may do without review.
D

Execution

Build, test, and self-correct within guardrails

Execution is where the agent moves fastest. It codes, tests, diagnoses, and corrects within approved boundaries, then submits deliverables for human review.

Working codeTest resultsSelf-correction logHuman sign-off
AgentBuilds the deliverable, runs tests, fixes issues, and produces work within approved guardrails.
HumanReviews against acceptance criteria, signs off, and escalates if guardrails are violated.
E

Verification

Evidence creation before release

Verification is not a checkbox. The agent executes tests, creates evidence, evaluates against criteria, and packages proof for human review.

Evidence packageCoverage reportConformance matrixRelease readiness verdict
AgentRuns self-tests, generates evidence artifacts, and summarizes verification results.
HumanReviews evidence, not just code, and approves the output for release readiness.
F

Operation

Deploy, monitor, and feed back into the cycle

Operation keeps ADLC alive after release. The agent watches production behavior, detects anomalies, and suggests rollbacks, patches, or flag changes while humans authorize release actions.

Production metricsAnomaly alertsRelease proposalsFeedback to Stage A
AgentMonitors production, identifies anomalies, and recommends safe operational actions.
HumanAuthorizes production changes, rollbacks, patches, and release decisions.

At a glance

Agent and human swimlane

The delegation pattern is simple: agents draft, generate, plan, build, test, and monitor; humans approve, accept, constrain, review, and authorize.

A
B
C
D
E
F
Agent
Drafts framings
Generates criteria
Creates task graph
Builds & self-corrects
Self-tests & evidence
Monitors & suggests
Human
Approves framing
Accepts spec
Sets guardrails
Reviews & signs off
Reviews evidence
Authorizes releases

Why this matters

The six stages turn agentic development into a governed operating model

🧭

Better framing

Work starts with clear intent, not vague prompts or open-ended tasks.

🚦

Controlled delegation

Agents gain execution room only after humans set boundaries and acceptance criteria.

📦

Evidence-driven release

Production decisions are based on verification evidence, not optimism.

Continue the ADLC series

Move from stages to gates

Use the series navigation and related resources to keep the ADLC pages connected inside the DataKnobs ecosystem.

PreviousAgent & Human Co-Development
CurrentADLC Six Stages
NextFive Gates
Agentic AIAI SDLCAI GovernanceGenerative AIData ProductsSlides