ADLC

The Agentic Development Lifecycle — a lifecycle built to handle uncertainty, prioritize behavioral guardrails, and treat deployment as the start of continuous learning rather than the end of a project.

The transition from SDLC to ADLC is the realization that agentic systems violate the core assumption of traditional software engineering: that behavior is fully specified at build time and strictly deterministic. Agents reason, adapt, and act across external tools. Small shifts in context compound into entirely different outputs.

Core shifts: SDLC vs ADLC

DimensionSDLCADLC
BehaviorDeterministic execution (same input = same output)Probabilistic execution (behavior evolves based on models, tools, and context)
Logic residencyCode and configHybrid of code, prompts, foundation models, and vector data
Success metricsFunctional pass/failAccuracy distribution, hallucination rates, token cost, and alignment
Feedback loopReactive (users report bugs, devs fix them)Recursive (telemetry feeds directly back into prompt context or auto-generates new intents)

The 8 stages

Commercially, Stages 0 through 1 through 2 through 3 encompass the Evidence Sprint.

StageNameCore question
0Preparation & HypothesesThe raw capture of intent and constraints. Not everything needs an agent.
1Scope Framing & Problem DefinitionEstablishing the human-agent boundaries and compliance limits.
2Agent Definition & ArchitectureDesigning the non-deterministic structure.
3Simulation & Proof of ValueThe commercial Evidence Gate. You do not pass this without external signal.
4Implementation & EvalsAgent-driven coding and baseline verification.
5TestingProbabilistic testing for reasoning, safety, and edge-cases.
6Agent Activation & DeploymentTiered release with strict autonomy guardrails.
7Continuous Learning & GovernanceManaging probabilistic drift and feeding the loop.

Stage 0 · Preparation & Hypotheses

The raw capture of intent and constraints. Not everything needs an agent.

Tactical practice

2–4 hour structured workshop. 10-minute silent assumption dump (Value, Usability, Feasibility, Viability, Operational). Failure premortem. No untagged claims survive.

Tooling

Meeting-transcription MCP (Whisper/Fathom); Miro/Slack MCP.

Agents & hooks
  • intent-elicitor — Probing agent that forces stakeholders to rephrase vague desires into testable assertions.
  • assumption-extractor — Parses live transcripts and auto-tags constraints.
  • premortem-adversary — Red-teams the room on AI edge-case failures.

Artifact: Raw intent.md, fully tagged. Facilitator validates tagging.

Stage 1 · Scope Framing & Problem Definition

Establishing the human-agent boundaries and compliance limits.

Tactical practice

Defining the agent's autonomy boundaries. Explicitly mapping what requires Human-in-the-Loop (HITL) approval. Drafting the initial evaluation criteria.

Agents & hooks
  • assumption-linter — Pre-commit hook. Rejects the commit of intent.md if any claim lacks an [ASSUMPTION: Category] tag or if autonomy boundaries are undefined.

Artifact: Refined intent.md with explicit human-agent responsibility mapping and operational KPIs.

Stage 2 · Agent Definition & Architecture

Designing the non-deterministic structure.

Tactical practice

Compressed design session defining the orchestration pattern, token economics (CAPEX/OPEX), and data architecture. Architecture must satisfy Stage 1 bounds without over-building.

Tooling

Figma MCP or repo-native prototype design via Claude Code.

Agents & hooks
  • .claude/skills/security-baseline — Organization-wide security compliance skill.
  • .claude/skills/ux-design-system — UX standards skill.
  • .claude/skills/adlc-agent-guardrails — Agent behavioral guardrails skill.
  • spec-compliance-linter — Verifying spec.md contains Agent Tool definitions, Failure Modes, Token Budget, Data Models.

Artifact: spec.md. Design/Tech Lead review.

Stage 3 · Simulation & Proof of Value

The commercial Evidence Gate. You do not pass this without external signal.

Tactical practice

Build an Evidence Slice or Vision Prototype. Attach it to a real client touchpoint. Product + Delivery decide: Promote / Iterate / Pivot / Stop. On a Promote, map scope blocks to the enterprise tracker and sync.

Tooling

Conversation Intelligence MCP (Gong/Zoom); Enterprise Tracker MCP (Jira/ADO).

Agents & hooks
  • synthetic-user-dryrun — Runs multi-persona simulation to test the agent before real client exposure.
  • evidence-synthesizer — Evaluates client signal strength, capping internal-only feedback at Directional.
  • gate-signoff-validator — Blocks push to /intent/ unless explicit decision tags are present.
  • tracker-sync — Automates mapping markdown scopes into enterprise tracker stories.

Artifact: intent.md (cleared) + per-project ESOA + optional Vision Prototype. PO merge sign-off.

Stage 4 · Implementation & Evals

Agent-driven coding and baseline verification.

Tactical practice

Git worktree isolation. Plan Mode first — Claude writes plan.md. Engineer accepts the plan, execution begins.

Tooling

Claude Code CLI; CLAUDE.md as the repo's living context engine.

Agents & hooks
  • .claude/agents/verifier.md — Runs the app and verifies behavior against plan.md.
  • .claude/agents/prompt-evaluator.md — Evaluates embedded AI outputs for hallucination/tool-calling accuracy during implementation.
  • path-blocking-hook — Intercepts file edits outside the scope defined in plan.md.

Artifact: plan.md and verified code diffs in isolated branches. Engineer plan acceptance.

Stage 5 · Testing

Probabilistic testing for reasoning, safety, and edge-cases.

Tactical practice

Non-interactive CI eval suite runs 20–50 task scenarios. Test files are read-only during bug-fix tasks — agents must fix code, not weaken the test.

Tooling

Headless Playwright/Puppeteer MCP; .github/workflows/agent-evals.yml.

Agents & hooks
  • lock-tests.sh — Blocks Edit/Write on tests/** during bug-fix runs.
  • eval-pass-checker — Blocks PR merge if task accuracy / safety alignment falls below the defined threshold.

Artifact: Verification logs, evaluation pass results. Automated CI pass threshold.

Stage 6 · Agent Activation & Deployment

Tiered release with strict autonomy guardrails.

Tactical practice

Multi-pass PR review (Bugs → Security → Alignment). Environment autonomy tiers: Dev = Autonomous, Staging = Clean CI + Auto-review, Prod = Named Human Sign-off.

Agents & hooks
  • .claude/agents/pr-reviewer.md — Multi-pass review outputting to REVIEW.md. Writes context updates back to CLAUDE.md if the agent repeats a mistake twice.
  • managed-settings-engine — permissions.deny, allowManagedHooksOnly, disableSideloadFlags.
  • network-egress-blocker — Hard network boundary enforcement preventing unauthorized agent tool-calls.

Artifact: REVIEW.md. Human Release Manager sign-off on Production.

Stage 7 · Continuous Learning & Governance

Managing probabilistic drift and feeding the loop.

Tactical practice

Statistical process control using Western Electric rules. When an agent hallucinates, breaches a boundary, or triggers a negative feedback loop in production, a background agent diagnoses it and writes a new intent back to Stage 0 automatically.

Tooling

Infra MCP (Datadog); Product Analytics MCP (PostHog/Pendo).

Agents & hooks
  • metric-watcher — Background monitor scanning bands.yaml for breaches.
  • incident-to-eval-compiler — Turns post-mortem logs into permanent regression tests in Stage 5.
  • Claude Tag — ChatOps bot allowing on-call engineers to tag the AI directly in incident threads.

Artifact: Updated bands.yaml, incident records, fresh intent.md pushed to Stage 0 triage queue.