Agent Teams Lifecycle
No company can outsource its way to digital excellence. Being digital means having your own bench of digital talent.
AI agents build governed. Humans ship trusted. 80% autonomy, 100% accountability.
Golden Path: Your First Agent Team
Phase 1: Assess (15 min)
Who: You (HITL) review capability gaps. meta-engineering-expert agent maps roles.
What: Evaluate your current talent gaps against the 15-agent model across 3 tiers.
Why: You cannot build a team without knowing what you lack. The 3-tier model (opus/sonnet/haiku) ensures the right cost-performance tradeoff per role.
What-if skip: Build without specialists; product-owner decisions go unchallenged; security gaps undetected.
How
bash scripts/governance-score.sh
Output
- Governance score (9 binary checks, target 100%)
- Capability gap report showing which agent roles are needed
- Tier mapping: Decision (opus) / Execution (sonnet) / Operations (haiku)
Quality Gate: Score visible. Gap analysis documented.
Phase 2: Compose (30 min)
Who: product-owner selects scope. cloud-architect maps architecture needs to agents.
What: Select agents for the project. Map them to Decision/Execution/Operations tiers.
Why: Agent selection must match project scope. Under-staffing wastes HITL time; over-staffing wastes tokens.
What-if skip: Wrong agents assigned, COORDINATION_WITHOUT_DELEGATION violations.
How
/speckit.plan # Generates agent delegation matrix
Output
- Agent roster with tier assignments
- Delegation matrix: which agent handles which domain
- Cost-performance tradeoff documented per agent
Quality Gate: PO + CA approve agent selection.
Phase 3: Coordinate (15 min)
Who: product-owner + cloud-architect establish authority chain.
What: Wire enforcement hooks. Establish PO+CA-first authority chain.
Why: Authority chain prevents STANDALONE_EXECUTION (anti-pattern #1). Hooks are deterministic.
What-if skip: STANDALONE_EXECUTION violations; ungoverned autonomous decisions.
How
Hooks are pre-configured in .claude/settings.json:
enforce-coordination.sh— blocks edits without PO+CA logsenforce-specialist-delegation.sh— blocks raw Edit/Write on domain files
Output
- Coordination hooks active (PreToolUse enforcement)
- Authority chain: PO first, CA second, specialists after
Quality Gate: bash scripts/governance-score.sh returns 100%.
Phase 4: Execute (variable)
Who: AI specialist agents build autonomously. HITL reviews evidence at Phase 3+ gates.
What: Specialist agents deliver within their domains via governed delegation.
Why: Governed execution: agents prepare, humans decide, humans commit. Each agent works in its domain.
What-if skip: Uncoordinated work, duplicate effort, NATO violations.
How
/speckit.implement # Processes tasks.md via specialist agents
Output
- Working software + test results
- Evidence files in
tmp/<project>/ - Coordination logs showing delegation chain
Quality Gate: Tests pass. Evidence on disk. No NATO violations.
Phase 5: Score (30 min)
Who: 3 scoring agents run sequentially (PO then CA then QA). HITL reviews consensus.
What: Sequential scoring round. Fix gaps first, then score once.
Why: Fix-then-score prevents SCORING_THEATER. Sequential prevents RACE_CONDITION_SCORING.
What-if skip: SCORING_THEATER — manufactured deltas from unchanged artifacts.
How
/speckit.retrospective # Triggers sequential PO+CA+QA scoring
Output
- Consensus score with percentage agreement
- Gap analysis with corrective actions
- Evidence of fix-then-score order
Quality Gate: >=95% agent agreement. Gaps have corrective actions with file paths.
Phase 6: Evolve (per sprint)
Who: meta-engineering-expert proposes. HITL decides what ships.
What: Upgrade agent frontmatter, retire underperformers, add new specialists.
Why: Agent frontmatter is model-version-agnostic. Upgrade the model, agents get smarter.
What-if skip: Stale definitions, capabilities plateau while models advance.
How
/speckit.improve # 6-phase improvement cycle with multi-agent consensus
Output
- Updated agent definitions in
.claude/agents/ - Improvement backlog items
- Anti-patterns extracted from incidents
Quality Gate: Improvement items tracked in .claude/improvement-backlog.md.
LEAN/5S Applied to Agent Teams
| Principle | Application | Evidence |
|---|---|---|
| Sort | 15 agents, not 50 — each has a clear domain | .claude/agents/ (15 files) |
| Set in Order | 3 tiers: Decision (opus) / Execution (sonnet) / Operations (haiku) | Agent frontmatter model: field |
| Shine | Governance score validates agent health | bash scripts/governance-score.sh |
| Standardize | Common frontmatter schema across all agents | name, description, tools, model |
| Sustain | /speckit.improve cycle evolves agents each sprint | Improvement backlog tracks changes |
By Persona
Solo Developer
Path: ln -s .adlc/.claude .claude → /speckit.specify → /speckit.implement
Time to Value: Full 15-agent team in under 5 minutes. No hiring, no ramp time.
Platform Team Lead
Path: git submodule add adlc-framework → governance-score.sh → /metrics:daily-standup
Time to Value: Team governance visibility in 1 day.
Enterprise Architect
Path: Review .claude/agents/*.md → /speckit.plan → /speckit.improve
Time to Value: Agent-capability mapping in 1 day.
Common Mistakes (Anti-Patterns)
| Mistake | Why It Fails | Fix |
|---|---|---|
| STANDALONE_EXECUTION | Specialist works without PO+CA approval | enforce-coordination.sh hook |
| SCORING_THEATER | Re-scoring unchanged artifacts | validate-rescore-freshness.sh hook |
| COORDINATION_WITHOUT_DELEGATION | PO+CA ran but Claude used raw Edit | enforce-specialist-delegation.sh hook |
| RACE_CONDITION_SCORING | Parallel scoring agents read different states | Sequential execution rule |
| AGENT_OVERPARALLELIZATION | 7 agents when 3 would suffice | Max 3 parallel agents rule |
| RUBBER_STAMP_COORDINATION | PO+CA in background then proceed | FOREGROUND-only rule |
Quick Reference: Command Cheat Sheet
# Assess agent team health
bash scripts/governance-score.sh
# Plan agent delegation
/speckit.plan
# Execute via specialists
/speckit.implement
# Score (sequential PO+CA+QA)
/speckit.retrospective
# Evolve agent definitions
/speckit.improve
# Daily team visibility
/metrics:daily-standup
Integration with Framework
| Component | Purpose | Location |
|---|---|---|
| 15 Agent definitions | Role, tools, model tier | .claude/agents/ |
| Coordination hooks | PO+CA enforcement | .claude/hooks/scripts/enforce-coordination.sh |
| Delegation hooks | Specialist domain enforcement | .claude/hooks/scripts/enforce-specialist-delegation.sh |
| Scoring freshness | Prevent stale re-scores | .claude/hooks/scripts/validate-rescore-freshness.sh |
| Governance score | 9-check agent health | scripts/governance-score.sh |
Last Updated: March 2026 | Status: Active | Maintenance: meta-engineering-expert