proffesor-for-testing__agentic-qe/docs/research/SUMMARY-task-agent-learning-gap.md

Executive Summary: Task Agent Learning Gap

Critical Finding: Claude Code's Task tool does NOT automatically capture agent output for learning persistence.

---

The Problem

When you spawn agents using Claude Code's Task tool:

Task("Generate tests", "Create test suite...", "qe-test-generator")

What happens:

• ✅ Agent executes successfully
• ✅ Output visible to user in conversation
• ❌ Learnings are NEVER persisted
• ❌ No automatic hook integration

---

Root Cause

Claude Flow's ReasoningBank Requires Trajectories

ReasoningBank exists but expects manual trajectory files:

{
  "steps": [
    { "action": "analyze", "result": "..." },
    { "action": "implement", "result": "..." }
  ],
  "query": "Task description",
  "metadata": { "agent": "test-gen" }
}

Problem: Nobody creates these trajectories for Task agents!

Hook System Limitation

Hooks work for Bash, Write, Edit but NOT for Task tool:

• Task agents run in separate threads
• No parent-child data flow
• Agent output stays internal to Claude Code

---

Current State Analysis

✅ What EXISTS in Claude Flow:

1. ReasoningBank - Learning system with judge + distill algorithms
2. Post-Task Hook - Script that processes trajectories
3. Memory Storage - SQLite database with embeddings
4. Retrieval API - Semantic search for memories

❌ What's MISSING:

1. Automatic trajectory capture from Task agents
2. Hook integration for Task tool
3. Structured output from agent conversations
4. Agent instrumentation for step tracking

---

Recommended Solution: Hybrid Self-Reporting

Architecture:

Agent Work
    ↓
Agent uses memory_store with "learning/*" key
    ↓
Background processor watches "learning/*"
    ↓
Converts to trajectory format
    ↓
Calls ReasoningBank judge + distill
    ↓
Persisted as reusable patterns

Implementation (3 parts):

1. Agent Learning API

// BaseAgent.ts
protected async recordLearning(learning: {
  type: 'pattern' | 'finding' | 'failure' | 'success',
  title: string,
  description: string,
  confidence: number
}): Promise<void> {
  await this.memory.store({
    key: `learning/${this.agentId}/${Date.now()}`,
    value: JSON.stringify(learning),
    namespace: "learnings",
    persist: true
  });
}

2. Updated Agent Instructions

## Learning Protocol

Record key learnings using:

mcp__agentic-qe__memory_store({
  key: "learning/[agent]/[timestamp]",
  value: JSON.stringify({
    type: "pattern",
    title: "Edge case handling",
    description: "Always validate null inputs...",
    confidence: 0.85
  }),
  namespace: "learnings",
  persist: true
});

3. Background Processor

# Process learnings periodically
aqe learn process

# Or automatic via cron
*/30 * * * * cd /project && aqe learn process

---

Why This Works

1. ✅ Works TODAY - No waiting for hook system changes
2. ✅ Explicit but simple - Clear protocol for agents
3. ✅ Integrates with MCP - Uses existing memory tools
4. ✅ Gradual rollout - Start with 1-2 agents, expand
5. ✅ Measurable - Can track learning coverage per agent

---

Timeline

Week 1: Foundation

• Implement BaseAgent.recordLearning()
• Update 2 agent instructions (test-gen, flaky-hunter)
• Test learning capture

Week 2: Automation

• Build LearningProcessor service
• Add CLI command aqe learn process
• Integrate with ReasoningBank

Week 3: Rollout

• Update all 19 QE agents
• Document learning protocol
• Monitor learning quality

---

Alternative Approaches (Not Recommended)

❌ Wait for Hook System Changes

• Timeline: Unknown, depends on Anthropic
• Risk: May never happen
• Blocker: Prevents all learning

❌ Parse Conversation History

• Challenge: Unstructured text parsing
• Accuracy: ~60-70% (lots of false positives)
• Maintenance: High (conversation format changes)

❌ Wrapper Scripts

• Complexity: High
• Adoption: Low (agents must use wrapper)
• Debugging: Difficult

---

Success Metrics

After implementing learning protocol:

1. Learning Coverage: % of agents recording learnings

   • Target: 80%+ within 1 month

2. Learning Quality: Avg confidence score of recorded patterns

   • Target: >0.75

3. Learning Reuse: % of tasks using retrieved memories

   • Target: 40%+ within 2 months

4. Time to Convergence: Iterations needed for pattern reuse

   • Target: <3 iterations per task type

---

Next Steps

1. Read: Full analysis in claude-flow-task-agent-learning-analysis.md
2. Decide: Approve hybrid self-reporting approach?
3. Build: Start with BaseAgent.recordLearning() implementation
4. Test: Pilot with qe-test-generator agent
5. Scale: Roll out to all agents

---

Key Takeaway

Claude Flow's ReasoningBank is powerful but requires intentional integration. The Task tool won't automatically persist learnings. We must build an explicit learning protocol that agents follow.

This is achievable and worth doing - the learning system will compound value over time.

---

Full Analysis: See claude-flow-task-agent-learning-analysis.md (12 sections, 800+ lines) Generated: 2025-12-08 Status: ✅ Research Complete - Ready for Implementation