Skip to content

Workflow Templates

Workflow templates are pre-defined "recipes" for common orchestration patterns. They are DAGs configured via YAML files which are converted into beads epics and tasks.


Built-in Templates

Template Description
Cook Sequential task execution with code review
Research to Tasks Research a topic and convert findings to actionable tasks
Debate Multi-agent debate for exploring solutions
Mediated Investigation Structured investigation with mediator
Research Proposal Collaborative proposal development
Quick Plan Rapid planning and task breakdown

List all available templates with:

perles workflows

Typical Workflow

A common development cycle uses multiple workflows in sequence:

1. Research Proposal

Start by generating a research document or proposal:

  • Launch the Research Proposal workflow
  • Provide a topic or problem description
  • The coordinator assigns workers to research and draft a proposal
  • Output: a proposal document in your configured document_path

2. Research to Tasks

Break down the proposal into actionable work:

  • Launch Research to Tasks with the proposal document
  • Workers analyze the proposal and create beads epics and tasks
  • Output: a structured epic with prioritized tasks

3. Cook

Execute the tasks with code review:

  • Launch Cook with the epic from step 2
  • The coordinator assigns tasks to workers sequentially
  • Each task goes through implementation, review, and commit phases
  • Workers automatically cycle through phases with built-in code review

Cook Workflow Details

Cook is the primary implementation workflow. It processes an existing epic's tasks in order:

  1. Task Assignment: Coordinator assigns the next task to an available worker
  2. Implementation: Worker implements the task (impl phase)
  3. Review: Worker reviews their own work or another worker reviews (review phase)
  4. Feedback: If review has issues, worker addresses feedback (feedback phase)
  5. Commit: Worker commits changes (commit phase)
  6. Next Task: Coordinator moves to the next task

Workers that run out of context are automatically replaced with fresh instances.