Tent.js

Start with

Light or dark

Option

output of agent

The type of the value that a run of the agent answers beside its text.

Value

A type: object({ category: choice(['bug', 'question', 'refund']), reason: text() }), or any other type. Left out, a run answers { text } alone.

With output the model is offered one more tool of the engine's own, answer, whose input is that type: a record type as it is, and any other type as { value }. The run ends when the model calls it with a value that fits, and answers { text, value }: the text the model wrote in that round, and the value. A value that does not fit is answered to the model as code.T0002, with a line for each field, as for any tool, and the model may call answer again. A run that ends with no call of answer fails with code.T7130: prose is never passed off as a value.

The value reaches whoever waits for the run: await sorter({ message }) in an action or a step answers { text, value }, evals holds it to expect.value, and the API and MCP answer it with the type in their descriptions. The out guards of the agent read the value too (option.agent.guards). A tool of the agent may not be told under the name answer (code.T1009).

Example

TypeScript
export const sorter = agent('sorter', {
  about: 'Sorts a message into a bug, a question or a refund',
  model: writer,
  instructions: 'Sort the message.',
  tools: [],
  access: 'signedIn',
  limits: { steps: 4, perRun: { tokensOut: 50000 } },
  output: object({ category: choice(['bug', 'question', 'refund']), reason: text() }),
})

const { value } = await sorter({ message: 'The page does not load.' })

Errors

code.T7130: the run ended with no call of answer.

code.T1009: a tool is told under the name answer.

See also

word.agent, option.agent.guards

Tent.js