Cortex · Prompt and context

Message order

A run’s prompt is put together once, when it starts, in this order: instructions, the tool guides, the output schema where one is sent, the finish protocol, the agent’s context, the run’s producers, then the input. The first four are the same on every run of an agent, so a provider’s prompt cache holds them.

Agent

const desk = defineAgent({
  id: 'desk',
  instructions: 'Help the member, in one sentence.',
  context: ['Doors open thirty minutes before the show.'],
  tools: [seatsLeft],  // with a guide
});

const preview = await desk.preview('How many seats are left on Saturday?', {
  llm,
  producers: ['The member asking is Grace.'],
});

Messages

system    Help the member, in one sentence.
system    TOOL GUIDES — how to use the tools on this run:
          ── seatsLeft ──
          Pass the night as YYYY-MM-DD.
          A night with no show answers left: 0.
system    FINISH PROTOCOL: To finish, call the `respond` tool
          ALONE in its own turn with the envelope {
          "reasoning"?: string, "response": string } as
          its arguments — or reply with ONLY that JSON
          envelope as your entire message.
system    Doors open thirty minutes before the show.
system    The member asking is Grace.
user      How many seats are left on Saturday?

Order

[
  'instructions',
  'tool guides',
  'finish protocol',
  'context',
  'producers',
  'input',
]

Output matches the expected value

Runs in this page on the installed package. The input and the expected output are defined by the example.