dag-planner

DAG Planner

Builds, validates, schedules, and dynamically modifies DAG execution graphs for complex multi-step tasks.

Decision Points

1. When to decompose a task into DAG nodes

Problem complexity assessment:
├── Single coherent output from one skill → No DAG needed (use direct skill)
├── 2-3 sequential steps, linear dependency → Simple pipeline (3-5 nodes)
├── Multiple parallel paths with convergence → Complex DAG (5-15 nodes)
└── Unclear requirements or multiple approaches → Start with vague nodes, refine later

2. Choosing node types by certainty level

Node type decision tree:
├── If task is well-defined with known skills → agent node
├── If task is unclear but bounded → vague node (refine during execution)  
├── If requires human judgment/approval → human-gate node
└── If task needs external system → agent node with tool restrictions

3. Determining granularity (critical decision)

Node size assessment:
├── If completable in 1 LLM call with 1-3 skills → Good granularity
├── If requires 5+ sequential LLM calls → Split into multiple nodes
├── If output is 1-2 simple values → Too granular, merge with parent
└── If node does "everything" → Too coarse, decompose further

4. Handling conflicts and dependencies

Dependency resolution:
├── Data dependency (B needs A's output) → Direct edge A→B
├── Resource conflict (both modify same file) → Serialize into different waves
├── Ordering constraint (B after A for logic) → Add ordering edge A→B
└── Optional dependency (B benefits from A but can run without) → Parallel with optional input

5. Dynamic modification triggers

Modification decision tree:
├── If node repeatedly fails → Replace node (different skill/model/approach)
├── If output quality insufficient → Add quality gate node after current
├── If gap discovered in coverage → Insert new node in dependency chain  
├── If path proves unnecessary → Remove node and reconnect edges
└── If multiple approaches needed → Fork into parallel paths, converge later

Failure Modes

Schema Drift

Symptoms: Downstream nodes fail with "missing field" or "unexpected format" errors Detection: If >50% of nodes in a wave fail with schema violations Diagnosis: Node output schemas don't match downstream input expectations Fix: Halt execution, audit all schemas in the dependency chain, add adapter nodes if needed

Circular Dependency Trap

Symptoms: Topological sort fails, no nodes have in_degree=0 after some are processed Detection: Validation throws CycleError during DAG build phase Diagnosis: Implicit dependency created a cycle (often through shared resources) Fix: Trace cycle using Kahn's algorithm state, break weakest dependency (usually resource conflicts)

Granularity Explosion

Symptoms: DAG has >20 nodes for simple task, execution overhead dominates work time Detection: If node_count > 3 × estimated_complexity_score Diagnosis: Over-decomposed trivial operations into separate nodes Fix: Merge adjacent nodes with single dependencies, combine related operations

Ghost Dependencies

Symptoms: Nodes wait indefinitely, execution stalls despite no explicit blocking Detection: If wave scheduling shows nodes stuck in higher waves despite dependencies met Diagnosis: Hidden dependencies through shared mutable state not modeled as edges Fix: Audit all file I/O and shared resources, add explicit edges or resource locks

Wave Starvation

Symptoms: Later waves have 1-2 nodes while early waves are overloaded Detection: If max_wave_size / min_wave_size > 5 and total_waves > 3 Diagnosis: Poor dependency structure creates artificial serialization Fix: Identify independent sub-paths, restructure to enable more parallelization

Worked Examples

Example 1: Simple Analysis Task

Problem: "Analyze customer feedback and generate recommendations"

Decomposition Process:

Assess complexity: Multiple data sources + analysis + synthesis → Medium complexity DAG
Initial nodes:
- collect_feedback (agent: data-collector)
- analyze_sentiment (agent: text-analyzer)
- generate_recommendations (agent: strategic-advisor)
Dependencies: collect → analyze → recommendations (linear)
Validation: No cycles, 3 waves
Expert insight: Novice would create single "analyze_everything" node; expert separates concerns

Final DAG:

nodes:
  - id: collect_feedback
    type: agent
    skills: [data-collector]
    output_schema: {feedback_items: array, metadata: object}
    wave: 1
  
  - id: analyze_sentiment  
    type: agent
    skills: [text-analyzer]
    depends_on: [collect_feedback]
    output_schema: {sentiment_scores: array, themes: array}
    wave: 2
    
  - id: generate_recommendations
    type: agent  
    skills: [strategic-advisor]
    depends_on: [analyze_sentiment]
    wave: 3

Example 2: Complex Research Task

Problem: "Research market trends, analyze competitors, and create strategy presentation"

Key Decisions:

Parallelization opportunity: Market research and competitor analysis are independent → parallel waves
Human gate needed: Strategy requires business judgment → human-gate node
Resource conflict: Both research tasks write to shared analysis.md → serialize or separate files

Decomposition reasoning:

Identify independent paths: market_research ∥ competitor_analysis
Both feed into strategy_synthesis
Human validation before final presentation
Expert catches: Presentation creation can start before human approval if we assume approval

Final structure: 4 waves, 2 parallel paths converging, 1 human gate

Example 3: Dynamic Replanning Scenario

Initial plan: Simple linear DAG for code review Runtime issue: Code quality gate fails repeatedly Replanning decision:

Diagnose: Quality threshold not met after 3 attempts
Fork approach: Add parallel path with different code analysis method
Convergence: Both paths feed into synthesis node that picks best result
New structure: Original linear becomes forked DAG with quality comparison

Expert insight: Novice would just retry same approach; expert recognizes need for alternative path

Quality Gates

Every node has explicit input_schema and output_schema defined
All dependencies are explicitly modeled as edges (no hidden shared state)
DAG passes cycle detection (topological sort completes successfully)
No node requires >3 skills or >5 minute estimated runtime
Wave distribution is reasonable (no wave has >3x nodes of smallest wave)
Resource conflicts are identified and resolved (files, APIs, external systems)
All vague nodes have refinement criteria specified
Human-gate nodes have clear approval criteria and timeout policies
Estimated total runtime is <10x single-agent equivalent approach
Each node's purpose can be explained in one sentence

NOT-FOR Boundaries

Do NOT use dag-planner for:

Single-step tasks: If only one skill needed → use skill directly
Pure sequential pipelines: If no parallelization possible → use simple workflow
Real-time execution: For running the DAG → use dag-runtime instead
Output validation: For checking node results → use dag-quality instead
Skill selection: For picking skills for nodes → use dag-skills-matcher instead
Data transformation: For ETL-style operations → use data pipeline tools
User interface flows: For UI workflows → use frontend orchestration tools

Delegate to:

dag-runtime: For executing planned DAGs
dag-quality: For validating node outputs
dag-skills-matcher: For selecting appropriate skills for each node
workflow-simple: For linear 2-3 step processes
Direct skill calls: For single-step tasks