Asking an agent

A workflow can hand a judgement to an agent: triage some notes, summarise a scan, pick a next step. It imports the agent and calls ask.

The files


clinic/
  index.ts      workflow
  triage.ts     agent

triage.ts:


import { chatAgent, model } from "@loops/sdk";
 
export const triageAgent = chatAgent({
  id: "triage",
  execute: async ({ submit, messages }) =>
    submit({
      messages,
      instructions:
        "You triage intake notes for a clinic. Decide whether the patient needs to be seen today.",
      model: model("gpt-4o-mini"),
    }),
});

Ask for a typed answer

Pass an output schema, and ask resolves to that shape:


import { Type, workflow } from "@loops/sdk";
 
import { triageAgent } from "./triage";
 
type Intake = { patientId: string; notes: string };
 
export const intakeWorkflow = workflow<Intake>({
  id: "intake",
  execute: async ({ step, event }) => {
    const verdict = await step.do("Triage the notes", () =>
      triageAgent.ask({
        message: event.body.notes,
        output: Type.Object({
          urgent: Type.Boolean(),
          reason: Type.String(),
        }),
      }),
    );
 
    return { patientId: event.body.patientId, ...verdict };
  },
});

The agent answers through the schema, so verdict is { urgent: boolean; reason: string }. If it doesn’t answer in that shape, the step throws and is retried.

Ask for the reply

Leave out output and ask resolves to the agent’s reply message:


const reply = await step.do("Summarise", () =>
  triageAgent.ask({ message: "Summarise these notes in one line: ..." }),
);
 
console.info(reply.content);

Inside a step

Call ask inside a step.do. Each ask starts a fresh conversation with the agent, named after the run and the step, so a retried step gets the answer it already had instead of asking the model again.

The agent has to be in the same project. Asking one in another project fails.