Agent Performance

Agent Performance Dashboard

Purpose

Provides visibility into agent usage patterns to optimize delegation and identify improvement opportunities.

When I Activate

I automatically load when you mention:

• "agent performance" or "agent metrics"
• "agent dashboard" or "agent usage"
• "which agents are used" or "underutilized agents"
• "agent success rate" or "agent statistics"

What I Do

1. Track Invocations: Record agent usage via workflow tracker
2. Measure Success: Track completion rates per agent
3. Analyze Patterns: Identify usage trends and gaps
4. Generate Reports: Create actionable dashboards

Quick Start

User: "Show me agent performance metrics"
Skill: *activates automatically*
       "Generating agent performance report..."

Core Capabilities

1. Report Generation

Generate a performance report by reading workflow logs and aggregating agent metrics:

User: "Generate agent performance report"

Report includes:

• Invocation counts per agent
• Success/failure rates
• Average completion times (when tracked)
• Underutilized agents list
• Recommendations for optimization

2. Live Tracking

Track agent invocations during workflow execution using the existing workflow_tracker:

# Already available in .claude/tools/amplihack/hooks/workflow_tracker.py
from workflow_tracker import log_agent_invocation

log_agent_invocation(
    agent_name="architect",
    purpose="Design authentication module",
    step_number=2
)

3. Metrics Storage

Metrics are stored in:

• Raw logs: .claude/runtime/logs/workflow_adherence/workflow_execution.jsonl
• Aggregated: .claude/runtime/metrics/agent_performance.yaml

Report Format

Summary Dashboard

# Agent Performance Summary
# Generated: 2025-11-25

total_invocations: 142

agents:
  architect:
    invocations: 45
    success_rate: 95.6%
    avg_duration_ms: 2340
    trend: increasing

  builder:
    invocations: 38
    success_rate: 89.5%
    avg_duration_ms: 4520
    trend: stable

  reviewer:
    invocations: 25
    success_rate: 100%
    avg_duration_ms: 1890
    trend: increasing

underutilized:
  - database (0 invocations in last 30 days)
  - integration (2 invocations in last 30 days)
  - patterns (3 invocations in last 30 days)

recommendations:
  - Consider using database agent for schema work
  - Integration agent available for external service connections
  - Patterns agent can identify reusable solutions

Implementation Guide

To Generate a Report

1. Read workflow execution logs:

   Read: .claude/runtime/logs/workflow_adherence/workflow_execution.jsonl
   

2. Filter for agent_invoked events:

   { "event": "agent_invoked", "agent": "architect", "purpose": "...", "step": 2 }
   

3. Aggregate by agent name:

   • Count invocations
   • Calculate success rates from workflow_end events
   • Compute average durations

4. Identify underutilized agents:

   • List all available agents from .claude/agents/amplihack/
   • Compare against invocation counts
   • Flag agents with <5 invocations in analysis period

5. Write report to:

   .claude/runtime/metrics/agent_performance.yaml
   

Available Agents Inventory

Core Agents (6):

• architect, builder, reviewer, tester, optimizer, api-designer

Specialized Agents (25):

• ambiguity, amplifier-cli-architect, analyzer, azure-kubernetes-expert
• ci-diagnostic-workflow, cleanup, database, documentation-writer
• fallback-cascade, fix-agent, integration, knowledge-archaeologist
• memory-manager, multi-agent-debate, n-version-validator, patterns
• philosophy-guardian, pre-commit-diagnostic, preference-reviewer
• prompt-writer, rust-programming-expert, security, visualization-architect
• worktree-manager, xpia-defense

Note: Agent count may change as specialized agents are added/removed. Use ls .claude/agents/amplihack/specialized/ for current count.

Tracking Best Practices

When Invoking Agents

Always log invocations for accurate tracking:

# Before invoking an agent via Task tool
log_agent_invocation(
    agent_name="security",
    purpose="Audit authentication implementation",
    step_number=7  # Optional: link to workflow step
)

# Then invoke the agent
Task(subagent_type="security", prompt="...")

Workflow Integration

The DEFAULT_WORKFLOW.md specifies agent delegation at each step. This skill helps verify adherence:

• Step 1: prompt-writer
• Step 2: architect
• Step 3: builder
• Step 4: tester
• Step 5: reviewer
• etc.

Configuration

Philosophy Alignment

This skill follows:

• Ruthless Simplicity: Uses existing infrastructure (workflow_tracker)
• Zero-BS: No placeholders, working aggregation logic
• Modular Design: Self-contained skill, clear boundaries
• Emergence: Insights emerge from simple tracking patterns

Interpreting Metrics

Success Rate Guidelines

Invocation Volume Interpretation

• High volume (30+ in 30 days): Core workflow agent, ensure reliability
• Medium volume (10-29): Regular use, monitor for optimization opportunities
• Low volume (5-9): Specialized use case, verify still needed
• Very low (<5): Consider if agent is discoverable or relevant

Duration Benchmarks

• < 2 seconds: Fast execution, typical for simple analysis
• 2-10 seconds: Normal for moderate complexity
• 10-60 seconds: Expected for deep analysis or multi-step tasks
• > 60 seconds: May indicate inefficiency, consider optimization

Empty State Handling

When no log data exists (new project or logs cleared):

# Agent Performance Report
# Period: Last 30 days
# Status: No data available

summary:
  total_invocations: 0
  message: "No agent invocations logged yet"

getting_started:
  - "Agent tracking begins when workflow_tracker logs invocations"
  - "Ensure agents are invoked via Task tool with proper logging"
  - "First report available after initial workflow execution"

next_steps:
  - "Run a workflow task to generate initial data"
  - "Verify workflow_tracker is properly configured"
  - "Check .claude/runtime/logs/ directory exists"

Limitations

This skill has the following constraints:

1. Depends on workflow_tracker: Only tracks agents invoked through the logging system
2. No real-time metrics: Reports are generated on-demand, not streamed
3. Historical data only: Cannot predict future usage patterns
4. Manual log analysis: Does not auto-detect anomalies or alert on issues
5. Single-project scope: Metrics are per-project, no cross-project aggregation
6. Time-based only: No correlation with code quality or PR outcomes