Evolving Workflow Skill

Evolving Workflow

Templates for building parallel AI agent workflows using the Codex SDK.

Templates

| Template | Pattern | Description | |----------|---------|-------------| | parallel_tasks | Fan-out | Process items in parallel with bounded concurrency | | multi_phase | Fan-out → Sequential → Fan-out | Three-phase workflow with intermediate planning |

Project Structure

your_workflow/
├── main.mbt           # Orchestration (usually unchanged)
├── moon.pkg.json
└── app/
    ├── job.mbt        # TaskHandle lifecycle (customize this)
    ├── args.mbt       # CLI parsing
    └── moon.pkg.json

main.mbt: Orchestration

The main file orchestrates: initialize → spawn parallel agents → summarize results.

async fn main {
  let config = @app.initialize() catch { ... }
  let codex = @codex.Codex::new()

  let results = @shared.for_all_tasks(
    config.items,
    async fn(idx, _) {
      let handle = config.task_start(idx)
      try {
        let thread = codex.start_thread(...)
        @async.retry(@async.RetryMethod::Immediate, fn() {
          let prompt = handle.prompt()
          let response = thread.run(prompt)
          handle.validate(response.final_response)
          handle.finish()
        }, max_retry=3)
      } catch { e => handle.error(e) }
    },
    parallelism=config.parallelism,
  )
  config.summarize(results)
}

For multi-phase, the pattern chains three phases: summarize (fan-out) → plan (sequential) → process (fan-out).

Customization

Edit app/job.mbt to customize:

| Function | What to change | |----------|----------------| | discover_items() | How to find targets to process | | read_item_content() | What data to load for prompts | | prompt() | Your task instructions | | validate() | Success criteria (raise to retry) | | finish() | How to save results |

Agent Skills: Evolving Workflow

Install this agent skill to your local

Skill Files