LIVE·6 Stages · Cost↓ · Risk↓ · Time-to-Value↑
ADLC · Framework Command Centre
The complete 6-stage product lifecycle — from first idea to retirement. Every stage has one owning command, one agent, and a signed gate artifact. Agents prepare. Humans decide at every gate.
WITH-WHAT · Framework command-centre (agents / commands / skills / hooks / governance) · Audience: CEO / CFO / CDO (lifecycle) · CTO / operator (authority chain + inner loop + DORA) · WITH-WHAT (/adlc) · HOW (/pm-toolkit) · Talent Bench →
§1 · AUTHORITY CHAIN (BLOCKING)
Source: CLAUDE.md "Authority Chain (BLOCKING)"
Agents Prepare. Humans Decide. Humans Commit.
Zero exceptions — not for agent confidence, not for bypass mode, not for verbal approval in chat. Principle I governs git mutations, terraform apply, and every irreversible production action.
👤
HITL
Human-in-the-Loop
Manager / T-Shape
All phases🎯
PO
Product Owner
Business validation
Phase 1–2🏗️
CA
Cloud Architect
Technical design
Phase 1–2⚙️
SPEC
Specialists
38 execution agents
Phase 3 [A]PHASE 1–2 [A] AUTONOMOUS
Think → Plan → Build. Agent-driven execution. PO+CA coordinate scope and design. Specialists execute under file-lock directives.
PHASE 3+ [HITL] APPROVAL REQUIRED
Test → Deploy → Monitor → Operate. HITL gates every phase transition. Agents prepare. Humans decide. Humans commit.
§2 · 6-STAGE PRODUCT LIFECYCLE
6 Stages · One Command Each
Every stage has a single owning command, a named agent, an autonomy tier, and a gate artifact the HITL must sign.
Documentationcross-cutting sub-step in every stage
Datameasurement in Stage 5 · migration in Stage 4
1
Discover
"Gate before any code — kill bad ideas early and cheaply."
Command
/adlc:discoverAgent
product-ownerWho decides
AI product-owner agent drafts the discovery brief; the Head of Platform (HITL) signs APPROVE / DEFER / KILL before any money is spent.
What the executive gets
A signed go/no-go discovery brief — bad ideas killed early and cheaply, before a single engineer is tasked.
2
Design
"Signed spec before first commit — no mid-build redesigns."
Command
/adlc:designAgent
cloud-architectWho decides
AI cloud-architect agent writes the module spec + low-level design; the Cloud Architect (HITL) signs off the spec with zero open clarifications.
What the executive gets
A signed technical spec + cost-shaped design — no mid-build redesigns, no surprise re-scoping.
3
Build
"Agents write and review code; humans approve the diff."
Command
/adlc, /speckit.implementAgent
code-reviewerWho decides
AI agents write and test the Terraform module; the code-reviewer agent screens it; the Head of Platform reviews and approves the diff.
What the executive gets
A registry-publishable module + green CI (60+ test gates passing) + updated docs — verifiable, not just claimed.
4
Deploy
"Human-approved release — nothing ships unreviewed."
Command
/commerce:tf-plan + tf-apply (HITL)Agent
infrastructure-engineerWho decides
Agent prepares the terraform plan + Infracost estimate + security scan; the CTO (HITL) approves and runs the apply — nothing ships unreviewed.
What the executive gets
A signed cost estimate + passing CIS/security scan + a human-approved release; post-apply cross-validation proves AWS matches the plan.
5
Support & Scale
"Agents monitor and surface; humans decide every remediation."
Command
/itsm:lifecycle, /ceremony:retroAgent
observability-engineerWho decides
AI observability agent monitors drift, DORA metrics, and incidents read-only and surfaces findings; the SRE / Head of Platform decides every remediation.
What the executive gets
A live reliability + DORA evidence pack and retro report; humans approve any change — no silent production edits.
6
Deprecate
"Nothing retires without a signed plan and evidence preserved."
Command
/adlc:deprecateAgent
lifecycle-stewardWho decides
AI lifecycle-steward agent drafts the decommission plan; the Head of Platform (HITL) signs it before anything is destroyed.
What the executive gets
A signed decommission plan with evidence preserved — clean, auditable retirement with no orphaned cost or compliance gap.
§3 · ENGINEERING INNER LOOP
Engineering Inner Loop — Nested, Not Competing
The constitution defines a 6-phase engineering loop: Plan → Build → Test → Deploy → Monitor → Operate. This is the inner loop nested inside Stage 3 (Build) and Stage 4 (Deploy) — not a separate product lifecycle.
5+1 Product Lifecycle (outer — CxO / boardroom vocabulary)
Discover
→Design
→Build
→Deploy
→Support & Scale
→Deprecate
Engineering Inner Loop (inside Build + Deploy — agent-engineering vocabulary)
Plan
→Build
→Test
→Deploy
→Monitor
→Operate
Experimentation Loop:
Build ↔ TestRuntime Optimization Loop:
Deploy → Monitor → OperateSource:
lifecycle-ssot/SKILL.md §3 · The "Operate" inner-phase retirement activity maps to Stage 6 Deprecate at the product level. Constitution ref: constitution.md:174§3.5 · ADLC STAGE-TO-COMMAND MAPPING (PI-004)
Framework generic · append ?product=<id> to filter by product
§4 · END-TO-END WALKTHROUGH
How to Use the Lifecycle End-to-End
Six commands in sequence. Each gate unlocks the next stage. HITL decision points shown in amber — agents cannot bypass them.
§5 · LIVE DORA METRICS + COMPONENT INVENTORY
Source: /data/dora.csv (live fetch) · /data/components.csv (live fetch) · Framework rows only
§6 · MEASURED ADOPTION
Counts sourced from /data/components.csv (live) + /data/dora.csv (live). No hardcoded forecasts.
3/sprint
DORA Band
target: ≥2/sprint
source: /data/dora.csv (live)
0%
Change Failure Rate
target: <5%
source: /data/dora.csv (live)
—
PDCA Composite
target ≥95%
source: /data/dora.csv (live)
§7 · KNOWN LIMITATIONS
Source: CLAUDE.md "Known Limitations" · cited: code.claude.com/docs/en/
Context degrades ~95%
Long sessions lose instruction adherence. Use /compact, delegate to subagents.
Subagent max depth = 1
Agents cannot spawn sub-agents. Orchestrator coordinates all specialist work.
Text output not hookable
Claude text responses bypass enforcement hooks. Rules-layer is the mitigation.
CLAUDE.md is advisory
Instructions shape behavior but don't enforce. Hooks provide deterministic enforcement for tool calls.
§8 · CITATIONS
Real quotes fetched from official Anthropic sources. No fabricated content.
"
The engineering question becomes how to cap the blast radius. There are broadly two ways to do this. The first is to supervise the agent's behavior via a human-in-the-loop.
Anthropic Engineering — How We Contain Claude (2026-05-25) →"
Claude Code is a command-line agent that runs in your terminal, reads your repository, edits files, executes commands, and requests confirmation before performing potentially destructive actions.
support.claude.com — Claude Code: Common Developer Use Cases →"
Opus 4.6 can complete software tasks which take humans almost 12 hours. When we first started building Claude Code, Sonnet 3.5 (new) was the frontier model and METR measured that it could do tasks that would take a human around 21 minutes. That's a roughly 41x jump in 16 months.
claude.com — Product Management on the AI Exponential (2026-03-19) →CxO Value at Every Stage
Cost ↓
Kill bad ideas at Discover before any dev spend. Gate artifacts prevent rework in Build.
Risk ↓
HITL gates at Discover, Deploy, and Deprecate ensure no unreviewed changes enter production or retirement.
Time-to-Value ↑
Autonomous Build + structured Design cuts cycle time. Signed specs mean no mid-sprint pivots.
Canonical SSOT:
.adlc/.claude/skills/governance/lifecycle-ssot/SKILL.md · ADR-020 (5-stage) + ADR-021 (Stage 6 Deprecate) · Talent Bench (38 agents + engineering discipline) → · PM Toolkit (HOW) →