Use a fixed workflow when the steps are known

How to do it

Write the task as steps. Some tasks always run the same steps, such as "classify the email, look up the order, draft a reply". Code those in order and call the model only inside steps that need it. A fixed workflow is cheaper, faster and easier to test than a loop. Keep the agent loop for open tasks, such as research across many sources. Give it the limits in this chapter and in limit cost, steps and loops.

Example

// A fixed workflow: the model works inside steps, code decides the order.
export async function handleSupportEmail(email: Email, user: User) {
  const { intent } = await ai.generate({ promptId: 'classify-email', vars: { body: email.body }, schema: Intent, userId: user.id });
  if (intent === 'return') {
    const order = await orders.latestFor(user.id);
    return ai.generate({ promptId: 'draft-return-reply', vars: { body: email.body, order: JSON.stringify(order) }, schema: Draft, userId: user.id });
  }
  return routeToPerson(email);
}

def handle_support_email(email: Email, user: User) -> Draft | None:
    intent = ai.generate("classify-email", {"body": email.body}, Intent, user.id).intent
    if intent == "return":
        order = orders.latest_for(user.id)
        return ai.generate("draft-return-reply", {"body": email.body, "order": order.model_dump_json()}, Draft, user.id)
    return route_to_person(email)

How to check

Each agent loop in the code has a decision note that lists the tasks it handles and why fixed steps could not. A search for agent loops finds none without a note.