Audit Agent Lifecycle

Purpose

Validates that agents can actually accomplish what they claim to do by examining the alignment between agent descriptions, prompt instructions, tool access, skill dependencies, and delegation patterns. This audit answers: given that all agents exist and their references resolve, can these agents execute their documented workflows correctly?

This is execution capability validation, not workflow coherence analysis.

When to Use

The model MUST activate this skill when:

• User requests agent capability validation
• User asks "can these agents do what they claim?"
• User wants to find dead agents or missing tool access
• User needs to validate agent-to-agent delegation contracts
• User wants to identify agents with contradictory instructions
• User needs to detect scriptable agent patterns (deterministic workflows)
• Post-refactoring validation of agent configurations
• Pre-marketplace review of agent quality

Relationship to Skill Lifecycle Audit

These two audits are complementary:

Aspect	skill-lifecycle-audit	agent-lifecycle-audit
Primary target	SKILL.md files	Agent .md files
Focus	Workflow coherence	Execution capability
Call graph direction	Skill → Skill, Skill → Agent	Agent → Agent, Agent → Skill
Contradiction scope	Cross-skill instructions	Cross-agent and agent-vs-skill

Run /audit-skill-lifecycle first to map workflow graphs, then this audit validates agents can execute those workflows. Both share the same patterns.md file in .claude/audits/.

Workflow

Step 1: Discovery

Scan plugin structure to identify:

• All agent files in agents/ directory
• Agent frontmatter configuration (tools, skills, model, disallowedTools, permissionMode)
• Skill references in agent prompts (Skill(), /skill-name patterns)
• Agent delegation patterns (Agent(subagent_type=), @agent-name references)
• Tool usage keywords in prompt body

Build dependency graph:

• Agent → Agent (delegation via Agent())
• Agent → Skill (loading via Skill() or skills field)
• Agent → Tool (usage via tool keywords in prompt)

Step 2: Analysis — Run Audit Dimensions

Execute 8 semantic audits across 3 depth tiers. See references/agent-lifecycle-audit.md for detailed dimension definitions.

Dimension	Tier	Detection Strategy	Output
1. Capability vs Configuration Alignment	2	Compare claimed purpose (description) against instructions (prompt) against available resources (tools/skills)	Per-agent capability matrix with misalignment flags
2. Skill Loading Correctness	2	Validate loaded skills provide expected capabilities and agent consumes outputs correctly	Per-agent skill load evaluation
3. Inter-Agent Contract Alignment	3	Verify delegating agents provide inputs/outputs matching target agent expectations	Contract alignment matrix between agent pairs
4. Prompt Instruction Contradictions	2	Detect internal, cross-agent, and agent-vs-skill contradictions	Contradiction pairs with file:line references
5. Tool Access Sufficiency	1	Map prompt instructions to required tools, check availability	Per-agent tool sufficiency report
6. Dead Agent Detection	1	Find agents never referenced by skills, commands, or other agents	List of potentially dead agents with evidence
7. Scriptable Agent Patterns	2	Identify deterministic workflows replaceable by scripts	List of scriptable patterns with steps
8. Self-Referential Pattern Learning	2	Classify issues by pattern, re-scan with new patterns until convergence	Append to shared patterns.md

Step 3: Capability Drift Check

Activate with Skill(skill: "plugin-creator:agent-capability-analyzer").

This phase compares each agent's static frontmatter description against its self-reported capabilities to detect description drift. Misrouted agents and stale descriptions are audit findings.

Single-agent audit (one agent being audited):

1. Read $CLAUDE_PLUGIN_ROOT/resources/describe-your-capabilities.template.md
2. Spawn one Task with subagent_type="<agent-id>", using the template as the prompt (replace AGENT_ID_HERE)
3. If the agent lacks Bash access, it returns capabilities as text — write it directly:

node $CLAUDE_PLUGIN_ROOT/scripts/update-agent-map.mjs --name "<agent-id>" --capabilities 'CAPABILITIES_TEXT'

4. Export the dataset:

node $CLAUDE_PLUGIN_ROOT/scripts/update-agent-map.mjs dump --file .claude/audits/agent-map.json

5. Perform inline gap analysis — compare description against capabilities in the exported JSON.

Full plugin audit (all agents):

1. Seed description fields from agent frontmatter:

node $CLAUDE_PLUGIN_ROOT/scripts/populate-agent-descriptions.mjs

2. Read $CLAUDE_PLUGIN_ROOT/resources/describe-your-capabilities.template.md
3. Spawn all agents simultaneously via Agent tool. Agents with Bash access write their own result to the DB. Collect text responses from agents without Bash access and write them via the update script.
4. After all Tasks complete, export the dataset:

node $CLAUDE_PLUGIN_ROOT/scripts/update-agent-map.mjs dump --file .claude/audits/agent-map.json

5. Perform gap analysis across all agents in the exported JSON.

Gap Analysis Categories

For each agent, identify entries in three categories:

• Underclaimed — in description, not in self-reported capabilities. Description may be aspirational or stale.
• Undocumented — in self-reported capabilities, not in description. Causes misrouting because the orchestrator cannot see these capabilities.
• Scope violations — self-reported capabilities outside the domain implied by the agent name and plugin context.

Gap Analysis Output Format

For each agent audited, produce a structured report:

Agent: <agent-id>
Description: <description field>
Capabilities: <capabilities field>

Underclaimed (in description, not in capabilities):
- <item>

Undocumented (in capabilities, not in description):
- <item>

Scope violations (outside expected domain):
- <item>

Verdict: ALIGNED | MINOR_DRIFT | SIGNIFICANT_DRIFT | SCOPE_VIOLATION

Use ALIGNED when all three lists are empty or contain only minor phrasing differences. Use SIGNIFICANT_DRIFT when Category 1 or 2 lists contain substantive functional differences. Use SCOPE_VIOLATION when Category 3 is non-empty.

Escalation

Agents with SIGNIFICANT_DRIFT or SCOPE_VIOLATION verdicts MUST be flagged as audit findings in the report. Pass these agents to subagent-refactorer for remediation:

Task is fixing description drift with subagent_type="plugin-creator:subagent-refactorer"
Context to include in the prompt: agents/<agent-name>.md (agent file), gap analysis report section for this agent
Output: revised agent file at agents/<agent-name>.md with corrected frontmatter description and rationale for each change

Step 4: Report Generation

Write audit artifacts to .claude/audits/:

1. agent-audit-report-{slug}.md — Full findings organized by dimension
2. agent-dependency-graph-{slug}.md — Visual graph of agent-to-agent and agent-to-skill delegation
3. patterns.md — Shared pattern catalog (appended to, shared with skill-lifecycle-audit)
4. agent-recommendations.md — Prioritized actionable fixes with severity levels

Audit Dimensions

Tier Strategy

Tier 1: Configuration Scan (fast, automated)

• Parse frontmatter for tools, skills, model, disallowedTools
• Extract Skill/Task/@ references from prompt
• Build agent dependency graph
• Check tool sufficiency against instruction keywords

Tier 2: Semantic Analysis (AI reasoning required)

• Read agent description and full prompt to understand claimed purpose
• Compare claimed purpose against actual instructions
• Evaluate whether skill loads make semantic sense
• Detect instruction contradictions with context awareness

Tier 3: Specialist Delegation (for specific issue types)

• Tool access ambiguity → delegate to agent that reads Claude Code tool inheritance docs
• Cross-agent contract mismatch → delegate to agent that evaluates input/output compatibility
• Prompt quality issues → delegate to @plugin-creator:subagent-refactorer for optimization (Agent tool: subagent_type="plugin-creator:subagent-refactorer")

1. Capability vs Configuration Alignment

Compare what the agent claims (description) against what it's instructed to do (prompt) against what resources it has (tools/skills).

Misalignment examples:

• Agent description says "runs tests" but no Bash tool access
• Agent prompt says "load the uv skill" but skills field doesn't include uv
• Agent prompt says "use WebFetch to check docs" but tool list doesn't include WebFetch

Output: Per-agent capability matrix showing CLAIMED, INSTRUCTED, AVAILABLE with flags where misaligned.

2. Skill Loading Correctness

For each skill reference in agent prompts:

• Is the skill reference namespace-qualified? (defer structural check to NamespaceReferenceValidator)
• Does the loaded skill actually provide the capability the agent expects?
• Does the agent pass appropriate inputs to the skill?
• Does the agent consume the skill's output format correctly?

Example issue: Agent loads /python3-development:modernpython and instructs "run modernpython to fix legacy patterns" but modernpython is a reference guide, not an automated fixer. The agent misrepresents what the skill does.

Output: Per-agent list of skill loads with semantic correctness evaluation.

3. Inter-Agent Contract Alignment

When agents delegate to other agents via Agent():

• Does the delegating agent's prompt describe inputs matching what the target agent expects?
• Does the delegating agent expect outputs in a format the target agent produces?
• Are there assumptions about shared state (files, directories, environment variables) not explicitly communicated?

Example issue: python-cli-architect delegates to python-pytest-architect with "create tests for this implementation." Does the target agent expect file paths? Architecture docs? Both? Does the delegating agent provide what's expected?

Output: Contract alignment matrix between agent pairs with delegation relationships.

4. Prompt Instruction Contradictions

Compare instructions within single agent prompt and across related agents:

Internal contradictions: Same agent says "always use Typer" in one section and "use argparse for CLI" in another, without conditional guards.

Cross-agent contradictions: Agent A says "format with ruff before linting" and Agent B says "lint first, then format" when both operate in the same workflow.

Skill-agent contradictions: Agent prompt gives instructions conflicting with the skill it loads (e.g., agent says "use unittest.mock" but loaded skill says "use pytest-mock exclusively").

Output: Contradiction pairs with file:line references, classification (internal/cross-agent/skill-agent), and whether guarded by conditions.

5. Tool Access Sufficiency

For each action the agent prompt describes:

• What tool is required to perform this action?
• Is that tool in the agent's tools field (or inherited)?
• If the tool is restricted via disallowedTools, is the agent trying to use it anyway?

Tool categories:

• Bash needed for: running commands, executing scripts, git operations
• Read/Write/Edit needed for: file operations
• Grep/Glob needed for: search operations
• Task needed for: delegating to other agents
• WebFetch/WebSearch needed for: documentation lookup
• MCP tools needed for: specific integrations

Output: Per-agent tool sufficiency report. Flag instructions requiring tools the agent doesn't have.

6. Dead Agent Detection

Identify agents that:

• Are registered in plugin.json but never referenced by any skill, command, or other agent
• Are referenced in skill documentation but not in any executable context (Skill(), Agent(), @agent)
• Have descriptions with trigger phrases that no workflow ever activates

Output: List of potentially dead agents with evidence (no inbound references found).

7. Scriptable Agent Patterns

Identify agents whose entire workflow could be replaced by a script:

• Agent prompt is a fixed sequence of tool calls with no conditional logic
• Agent doesn't need AI reasoning — just runs commands in order
• Agent's "decisions" are deterministic based on file existence or command output

These are candidates for conversion to Python scripts or commands, which execute faster, cost nothing, and are deterministic.

Output: List of scriptable agent patterns with the deterministic steps identified.

8. Self-Referential Pattern Learning

Same approach as skill-lifecycle-audit dimension 7:

1. Classify each discovered issue by detection pattern
2. Record the pattern with description, heuristic, and example
3. Re-scan all already-audited agents using the newly discovered pattern
4. Repeat until convergence

Output: Append to the shared patterns.md file (same file as skill-lifecycle-audit).

Output Format

agent-audit-report-{slug}.md

# Agent Lifecycle Audit Report: {plugin-name}

**Audit Date:** {ISO timestamp}
**Plugin Path:** {absolute path}
**Agents Audited:** {count}

---

## Summary

| Dimension | Issues Found | Severity Breakdown |
|-----------|-------------|--------------------|
| Capability vs Configuration Alignment | {count} | Error: {n}, Warning: {n}, Info: {n} |
| Skill Loading Correctness | {count} | Error: {n}, Warning: {n}, Info: {n} |
| Inter-Agent Contract Alignment | {count} | Error: {n}, Warning: {n}, Info: {n} |
| Prompt Instruction Contradictions | {count} | Error: {n}, Warning: {n}, Info: {n} |
| Tool Access Sufficiency | {count} | Error: {n}, Warning: {n}, Info: {n} |
| Dead Agent Detection | {count} | Info: {n} |
| Scriptable Agent Patterns | {count} | Info: {n} |
| Capability Drift Check | {count} | Error: {n}, Warning: {n} |

---

## Dimension 1: Capability vs Configuration Alignment

### Agent: {name}

**Claimed capability (description):**
{excerpt from description field}

**Instructed actions (prompt):**
{excerpt from prompt body}

**Available resources:**
- Tools: {tools field value}
- Skills: {skills field value}

**Finding:**
❌ **ERROR** Agent claims to "run tests" but has no Bash tool access
{file}:{line}

**Severity:** error (agent cannot accomplish its purpose)

---

{repeat for each dimension}

---

## Recommendations

See `agent-recommendations.md` for prioritized actionable fixes.

agent-dependency-graph-{slug}.md

# Agent Dependency Graph: {plugin-name}

## Agent → Agent Delegation

```mermaid
graph TD
    A[python-cli-architect] -->|Task| B[python-pytest-architect]
    A -->|Task| C[python-code-reviewer]
    B -->|Task| D[stdlib-scripting skill]

Agent → Skill Loading

graph LR
    A[python-cli-architect] -->|loads| S1[/python3-development:modernpython/]
    A -->|loads| S2[/python3-development:uv/]
    B[python-pytest-architect] -->|loads| S3[/python3-development:testing/]

Tool Usage Matrix

Agent	Bash	Read	Write	Edit	Grep	Glob	Task	WebFetch
python-cli-architect	✅	✅	✅	✅	✅	✅	✅	❌
python-pytest-architect	✅	✅	✅	✅	❌	❌	❌	❌

Isolation Analysis

Agents with no inbound references (potentially dead):

• {agent-name} — no skills, commands, or agents reference this agent

Agents with no outbound references (terminal nodes):

• {agent-name} — does not delegate to other agents or load skills

### patterns.md (shared with skill-lifecycle-audit)

```markdown
# Shared Pattern Catalog

**Last Updated:** {ISO timestamp}
**Total Patterns:** {count}

---

## Pattern: Missing Tool Access for Workflow Step

**Detection Heuristic:**
1. Search prompt for action keywords: "run", "execute", "install", "test", "validate"
2. Map each keyword to required tool (Bash for command execution)
3. Check if tool is in tools field or inherited
4. Flag if tool missing

**Example Finding:**
Agent `test-runner` has instruction "Run pytest tests" but tools field is `["Read", "Write"]` — missing Bash.

**Source Audit:** agent-lifecycle-audit dimension 5 (Tool Access Sufficiency)

**Found In:**
- plugin-creator:refactor-executor (line 42)
- python3-development:python-pytest-architect (line 67)

---

{repeat for each pattern}

Additional Resources

• references/agent-lifecycle-audit.md — detailed audit dimension definitions, detection strategies, and severity criteria
• ../audit-skill-lifecycle/SKILL.md — complementary audit for workflow coherence