Skip to main content

Autonomous Planning

Autonomous planning is Radium's AI-powered system for generating executable plans from natural language specifications. It automatically structures your requirements into iterations and tasks, validates dependencies, and creates execution workflows.

Overview​

The autonomous planning pipeline consists of four main stages:

  1. Plan Generation: AI generates a structured plan from your specification
  2. Validation: Multi-stage validation ensures plan correctness
  3. Dependency Analysis: DAG construction and cycle detection
  4. Workflow Generation: Creates executable workflow from validated plan

Plan Generation​

Generate a plan from a specification using the rad plan command:

# Generate plan from direct input
rad plan "Build a REST API with authentication"

# Generate plan from file
rad plan spec.md

# Generate plan with custom ID
rad plan --id REQ-123 spec.md

The plan generator uses AI to extract:

  • Project name and description
  • Tech stack requirements
  • Iterations with goals
  • Tasks with dependencies
  • Agent assignments
  • Acceptance criteria

Validation Pipeline​

The autonomous planner performs multi-stage validation:

Stage 1: Dependency Graph Validation​

Validates that all task dependencies exist and detects circular dependencies:

# Plan with valid dependencies
rad plan "Task 1 depends on nothing, Task 2 depends on Task 1"

# Plan with circular dependency (will fail validation)
rad plan "Task 1 depends on Task 2, Task 2 depends on Task 1"

Stage 2: Agent Assignment Validation​

Verifies that assigned agents exist in the agent registry:

# Valid agent assignment
rad plan "Use code-agent to implement feature"

# Unknown agent (warning, not error)
rad plan "Use unknown-agent to implement feature"

Stage 3: Dependency Reference Validation​

Ensures all dependency references use valid task ID format (I[number].T[number]):

# Valid dependency reference
rad plan "Task 2 depends on I1.T1"

# Invalid dependency reference (will fail)
rad plan "Task 2 depends on I5.T1"  # I5 doesn't exist

Validation Retry Logic​

If validation fails, the planner automatically retries up to 2 times with validation feedback:

  1. Initial Generation: AI generates plan from specification
  2. Validation Check: Plan is validated for correctness
  3. Retry with Feedback: If validation fails, feedback is provided to AI for regeneration
  4. Final Validation: Regenerated plan is validated again

This retry mechanism helps the AI correct common issues like:

  • Missing dependencies
  • Invalid task references
  • Circular dependencies

Example: Complete Planning Workflow​

# 1. Create specification file
cat > my-spec.md << EOF
# My Project

Build a web application with user authentication.

## Iteration 1: Setup

1. **Task 1** - Setup project structure
   - Agent: code-agent
   - Dependencies: 
   - Acceptance Criteria:
     - Project structure created
     - Dependencies installed

2. **Task 2** - Implement authentication
   - Agent: code-agent
   - Dependencies: I1.T1
   - Acceptance Criteria:
     - Login endpoint working
     - JWT tokens generated
EOF

# 2. Generate plan
rad plan my-spec.md

# 3. Plan is validated automatically
# 4. If validation fails, retry with feedback
# 5. Validated plan is saved to .radium/plan/REQ-XXX/

Validation Errors​

Common validation errors and how to fix them:

Circular Dependency​

Error: Circular dependency detected: I1.T1 -> I1.T2 -> I1.T3 -> I1.T1

Fix: Remove or reorder dependencies to break the cycle.

Missing Dependency​

Error: Task I1.T2 references non-existent dependency: I5.T1

Fix: Ensure the referenced task exists in the plan, or remove the invalid dependency.

Invalid Task ID Format​

Error: Invalid task ID format: INVALID

Fix: Use format I[number].T[number] where number is the iteration/task number.

Best Practices​

  1. Start Simple: Begin with a single iteration and few tasks
  2. Use Clear Dependencies: Explicitly state task dependencies
  3. Assign Agents: Specify which agent should handle each task
  4. Define Acceptance Criteria: Clear criteria help validation and execution
  5. Iterate: Use validation feedback to refine your plan

Troubleshooting​

Plan Generation Fails​

  • Check that your specification is clear and well-formatted
  • Ensure you have a valid workspace (rad init if needed)
  • Verify AI model access and credentials

Validation Always Fails​

  • Review validation error messages carefully
  • Check for circular dependencies
  • Verify all dependency references use correct format
  • Ensure assigned agents exist in your workspace

Retry Logic Not Working​

  • Validation retries are automatic (up to 2 retries)
  • If validation fails after retries, fix the specification manually
  • Check that your AI model supports the planning prompt format

See Also​