Build & Deploy AI Agents

Exercise 3

Build: Workflows

Fan out to parallel specialists with risk tiers and a coordinator.

Build Module 3 on top of your Module 2 work (or checkpoint-2-tools-ai).

The goal

Turn a single review call into a durable pipeline:

  • a ReviewWorkflow with checkpointed steps,
  • assessRisk that picks a tier and the specialists to run,
  • specialists running in parallel with Promise.all,
  • a coordinator pass that dedupes and decides the verdict, then writes back to state.

Getting started

Let your agent scaffold the WorkflowEntrypoint. Keep each step inside its own step.do() so it is checkpointed and retried.

Critical gotcha

You must re-export ReviewWorkflow from server.ts and bind it in wrangler.jsonc. If you forget, Cloudflare does not error, the workflow just never runs.

Stuck?

git checkout checkpoint-3-workflows -- src/

Done when

  • A review on a multi-file PR runs more than one specialist and the findings panel fills in when it finishes.
  • You can explain why step.do() checkpointing matters.