Agent Planning Data

Structured Planning Data

Structuring complex task planning into directed acyclic graphs (DAGs) through human annotation, generating high-quality training data for planner models' task decomposition and multi-agent coordination capabilities.

Demo Cases

Interactive DAG visualizations from our annotated dataset

Query

Build a full-stack todo web application with authentication, database, and deployment.

Click any node to view detailed annotation

Understand Require…Architecture DesignDatabase SchemaAuth ModuleAPI RoutesFrontend UIIntegrationTestingDeployment
StartPlanningExecutionVerificationTerminal
FromToData Flow
rootarchparsed_requirements
archdbtech_stack_config
archauthauth_requirements
archapiapi_spec
dbfeschema_types
authfeauth_context
dbintegratedb_client
authintegrateauth_middleware
apiintegrateroute_handlers
fetestcomponents
integratetestapi_endpoints
testdeploytest_results

From Todo Lists to DAGs

Why linear planning tools are no longer enough

Linear Todo List

  • Models need frequent reminders to stay on track (system reminder every 5 turns)
  • Fixed checklists limit the model's ability to autonomously adjust
  • Sub-agents cannot share coordination state
  • Linear structure cannot express parallelism or dependencies

DAG Task Graph

  • Dependencies between tasks are explicitly modeled (directed edges)
  • Supports parallel execution of independent subtasks
  • Agents can pass context and status updates through edges
  • Models can dynamically add or remove nodes, adjusting plans flexibly

Reference: Anthropic's Claude Code team found that as model capabilities improved, the linear TodoWrite tool became a bottleneck. They introduced the Task Tool with dependency support and cross-agent communication (DAG structure), significantly improving multi-agent collaboration.

View original post

What We Annotate

01

DAG Topology

Complete graph structures with nodes (subtasks) and directed edges (dependencies)

02

Node Metadata

Agent roles, tool invocations, input/output descriptions for each node

03

Execution Semantics

Parallel groups, critical paths, data flow direction, and error handling strategies

Annotation Workflow

1

Task Collection & Design

  • Collect real user queries
  • Filter for complex tasks suitable for DAG decomposition (exclude single-step tasks)
  • Design annotation templates: node fields + edge fields + metadata
2

DAG Structure Annotation

  • Annotators decompose queries into subtask nodes (with agent roles, tools, descriptions)
  • Annotate inter-node dependencies (directed edges), ensuring acyclicity
  • Annotate parallelism: which nodes can execute concurrently
  • Annotate data flow direction for each edge (input/output mapping)
3

Quality Review

  • Cross-review: another annotator independently reviews DAG structural validity
  • Automated validation: cycle detection, connectivity checks, orphan node detection
  • Consistency assessment: compare two annotators' DAGs, compute structural similarity
  • Expert arbitration: senior annotator resolves disagreements
4

Data Augmentation & Training

  • Convert annotated DAGs to JSON format training data
  • Data augmentation: generate multiple valid DAG variants for the same query
  • Train Planner models: input query → output DAG structure
  • Evaluation: compare with ground-truth DAGs for topological similarity & execution success rate

Annotation Guidelines

Node Granularity

Each node should correspond to an independently completable subtask, roughly equivalent to one agent tool call

Edge Semantics

Edges represent data/control dependencies, must annotate data flow direction, no redundant edges allowed

Parallel Annotation

Independent nodes should be marked as parallelizable and explicitly declared in parallel_groups

Agent Assignment

Each node specifies an appropriate agent role and available tool set, ensuring executability

Data Format Example

{
  "query": "Build a full-stack todo app with auth and deploy",
  "dag": {
    "nodes": [
      { "id": "arch",  "label": "Architecture Design",  "agent": "Architect",  "tools": ["SequentialThinking"] },
      { "id": "db",    "label": "Database Design", "agent": "DB-Agent",   "tools": ["Bash","Write"] },
      { "id": "auth",  "label": "Auth Module",   "agent": "Auth-Agent", "tools": ["Write","Read"] }
    ],
    "edges": [
      { "from": "arch", "to": "db",   "data_flow": "tech_stack_config" },
      { "from": "arch", "to": "auth", "data_flow": "auth_requirements" }
    ],
    "parallel_groups": [["db", "auth", "api"]],
    "critical_path": ["arch", "db", "fe", "test", "deploy"]
  },
  "metadata": { "annotator": "expert_03", "time_spent_min": 11, "difficulty": "medium" }
}

Annotation Quality Metrics

92%
Inter-Annotator Agreement
100%
Acyclicity Validation Pass Rate
50+
Covered Task Scenarios
2,000+
Annotated DAG Samples