Skip to content

SELF_DIRECTED_DECOMPOSITION.md

Latest commit

 

History

History
129 lines (102 loc) · 4.37 KB

SELF_DIRECTED_DECOMPOSITION.md

File metadata and controls

129 lines (102 loc) · 4.37 KB

Self-Directed Task Decomposition - IMPLEMENTED ✅

What Changed

Before (Hardcoded Limits)

  • Single-shot: AI generated graph → STOP (no continuation)
  • Or: Fixed 2-5 iterations regardless of complexity
  • Fixed node counts: 8-15 nodes hardcoded
  • No autonomy: System decides when to stop

After (True Agent Autonomy)

  • Self-directed phases: AI decides how many phases needed
  • Autonomous evaluation: After each phase, AI reviews and decides
  • No iteration limits: Only safety limits (50 phases, 200 nodes for extreme edge cases)
  • Full context: AI sees ALL nodes (not truncated to 10)
  • Explicit completion: AI signals "complete" when comprehensive

How It Works

Flow

User: "Create Greek mythology graph"

Phase 1:
  AI: Generates 12 Olympians
  System: Commits to store
  System: "Phase complete. Evaluating next phase..."
  
Phase Evaluation:
  AI receives: Full graph state (all 12 nodes listed)
  AI evaluates: "Main Olympians complete. Need to add Titans for generational context."
  AI decision: "continue"
  AI generates: 8 Titans
  
Phase 2:
  System: Commits 8 Titans
  System: "Phase complete. Evaluating next phase..."
  
Phase Evaluation:
  AI receives: Full graph state (all 20 nodes: Olympians + Titans)
  AI evaluates: "Graph now comprehensive with major deities. Hierarchies established."
  AI decision: "complete"
  
Result: 20-node graph created in 2 autonomous phases

Decision Schema

// Continue (needs more work)
{
  "decision": "continue",
  "reasoning": "Main Olympians complete (12 nodes). Now adding 8 Titans to show generational hierarchy.",
  "response": "Adding Titans to expand the pantheon...",
  "graphSpec": {
    "nodes": [...8 Titan nodes...],
    "edges": [...relationships...]
  }
}

// Complete (graph is comprehensive)
{
  "decision": "complete",
  "reasoning": "Graph now has 30 Greek deities covering Olympians, Titans, and Heroes with family relationships.",
  "response": "✅ Greek mythology graph complete with 30 deities!"
}

Files Modified

1. bridge-daemon.js

Lines 338-351: Updated initial planner to explain self-directed execution

  • Removed "5-iteration budget" language
  • Added examples of autonomous phasing
  • Emphasized starting with substantial first phase

Lines 431-439: Updated initial phase requirements

  • Changed from fixed node counts to flexible foundation
  • Emphasized core concepts over exhaustiveness

Lines 924-953: Removed iteration limits, added safety checks

  • Removed MAX_ITERATIONS = 2
  • Removed REASONABLE_NODE_COUNT = 15
  • Added extreme safety limits: 50 phases, 200 nodes
  • Changed from warn to debug for phase logging

Lines 1156-1212: Complete rewrite of continuation prompt

  • Changed from "ITERATION X/Y" to "PHASE EVALUATION"
  • Provide ALL nodes (not truncated to 10)
  • Clear "continue" vs "complete" options
  • Evaluation guidelines based on topic complexity
  • Emphasis on being decisive and comprehensive

2. src/services/Committer.js

Lines 329-388: Re-enabled agent-controlled continuation loop

  • Restored continuation call to /api/ai/agent/continue
  • Send full graph state with ALL nodes
  • Pass originalMessage and conversationHistory
  • AI decides next action autonomously

Safety Measures

  1. Phase limit: 50 phases max (extreme edge case)
  2. Node limit: 200 nodes max (sanity check)
  3. Timeout: Existing request timeouts still apply
  4. Error handling: Graceful fallback if AI fails to respond

Expected Behavior

Simple Topics (1 phase)

  • "Solar system" → 9 nodes → COMPLETE
  • "Days of week" → 7 nodes → COMPLETE

Medium Topics (2-3 phases)

  • "Greek mythology" → Phase 1: 12 Olympians → Phase 2: 8 Titans → COMPLETE (20 nodes)
  • "US Presidents" → Phase 1: Founding era → Phase 2: Modern era → COMPLETE

Complex Topics (4-6 phases)

  • "World War II" → Phase 1: Major powers → Phase 2: Leaders → Phase 3: Key battles → Phase 4: Outcomes → COMPLETE (40-60 nodes)

Testing

Ready to test with real requests. Expected improvements:

  • ✅ Faster for simple topics (1 phase vs 5 iterations)
  • ✅ More comprehensive for complex topics (AI decides when sufficient)
  • ✅ No arbitrary stopping (AI evaluates, not hardcoded limits)
  • ✅ Better quality (AI sees full context, makes informed decisions)

Bridge Restarted

The bridge daemon has been restarted with the new autonomous agent loop active.