Streaming answers

An agent can answer POST /invoke with a stream instead of JSON: a Vercel AI SDK UI message stream, AG-UI, server-sent events, NDJSON, plain text — any format. A caller who started the run with ?passthrough=true receives it byte for byte as you write it, so a chat UI can read your agent’s own stream format with no mapping in between.

How AgentOS tells JSON from a stream

An answer whose Content-Type is application/json (or that has none) is the contract’s JSON.
Any other Content-Type is a stream. AgentOS doesn’t read into it; it relays it to a passthrough caller.

Because AgentOS doesn’t read your stream, your agent reports the run’s outcome itself, before the stream ends.

Report the result

Post the same JSON you would have answered, with the run token:

POST {api_url}/v0/runs/{run_id}/result
x-agentos-run-token: <context.token>

{"status": "completed", "output": {"text": "..."}}

It answers 202 {"status": "accepted"}. The body is one of:

{"status": "completed", "output": {...}} — the output is checked against the capability’s output schema at once; a mismatch answers 422 invalid_output with errors, so you can correct it and post again;
{"status": "failed", "error": {"code": "...", "message": "..."}};
{"status": "suspended", "hold_id": "h_7"} — after asking a person.

A result is at most 1 MiB and is accepted once per invocation: the same result again is accepted too, a different one is 409 result_already_reported. It is scrubbed of secrets like a JSON answer, and recorded when the stream ends.

Post the result before your stream ends, or at the latest within 3 seconds of its end. A later one is refused with 401: the invocation’s token has ended.

With the SDKs

Return a stream and set the output; the SDK posts the result for you before the stream ends:

import agentos

def tell(inv: agentos.Invocation):
    def pieces():
        text = ""
        for delta in inv.llm.stream([{"role": "user", "content": inv.input["topic"]}]):
            text += delta
            yield delta
        inv.set_output({"text": text})          # reported as the run's output

    return agentos.Stream(pieces(), "text/plain; charset=utf-8")

agentos.serve({"tell": tell}, port=8080)

import { serve, stream } from "@agentos/sdk";

await serve({
  tell(inv) {
    return stream((async function* () {
      let text = "";
      for await (const delta of inv.llm.stream([{ role: "user", content: inv.input.topic }])) {
        text += delta;
        yield delta;
      }
      inv.setOutput({ text });                   // reported as the run's output
    })(), { contentType: "text/plain; charset=utf-8" });
  },
}, { port: 8080 });

If the platform refuses the result (an output that doesn’t match the schema), the SDK reports failed instead, so a run never completes silently with a bad output. A caller that goes away ends the stream and fails the run.

Without a result

If no result arrives, AgentOS waits those 3 seconds after the stream ends, then decides:

A 2xx body of at most 1 MiB that is the contract’s JSON ({"status": "completed" | "failed" | "suspended", ...}) but sent with another Content-Type — some frameworks send JSON as text/html — is read as the contract, at once.
A stream that ended cleanly with a 2xx status holds the run if this invocation asked a person, and otherwise completes it with output: {"streamed": true, "bytes": ..., "content_type": "..."}. That output isn’t checked against the output schema, and the run page marks the run “streamed”.
Anything else fails the run with agent_interrupted, a non-2xx error page included.

Limits

A stream may pause up to the capability’s timeout between writes.
A stream may be at most 10 MiB; more fails the run with invalid_response.

Asking a person mid-stream

To ask a person in the middle of a stream: create the hold, end your stream the way your format ends it, and report {"status": "suspended", "hold_id": "h_7"}. The resume that follows the decision isn’t streamed to anyone; answer it with JSON, or with a stream and a result, and callers read its output from the run.

Never stream secrets

What you stream reaches the caller exactly as you wrote it. Unlike the output, error and events AgentOS stores, a stream is not scrubbed of secrets. Never write context.token, a secret or the invoke body into it.

Reading the stream as a caller

Start the run with ?passthrough=true. The response is exactly what the agent answered — its status code, Content-Type, and body as it arrives — plus an X-AgentOS-Run-Id header:

curl -sS -N -D - -X POST "$AGENTOS/v0/envs/default/agents/stream-agent/capabilities/chat/runs?passthrough=true" \
  -H "Authorization: Bearer $AGENTOS_API_KEY" -H "Content-Type: application/json" \
  -d '{"input": {"message": "What is AgentOS?"}}'

A chat app’s backend calls this with its API key and hands the response to the browser unchanged, so a UI built on the AI SDK’s useChat, for example, reads it directly. Never put an API key in a browser. Closing the connection doesn’t stop the run; cancel it. After the stream, GET the run for its outcome. See the Runs API for headers, replays and limits.

examples/stream-agent streams a model’s answer as a Vercel AI SDK UI message stream, asks a person mid-stream, and turns a guardrail block into a clean failed run. See Examples.

GET STARTED

BUILD AGENTS

SDKS & TOOLS

FRAMEWORK EXAMPLES

PLATFORM

API REFERENCE

How AgentOS tells JSON from a stream

Report the result

With the SDKs

Without a result

Limits

Asking a person mid-stream

Never stream secrets

Reading the stream as a caller

​How AgentOS tells JSON from a stream

​Report the result

​With the SDKs

​Without a result

​Limits

​Asking a person mid-stream

​Never stream secrets

​Reading the stream as a caller

How AgentOS tells JSON from a stream

Report the result

With the SDKs

Without a result

Limits

Asking a person mid-stream

Never stream secrets

Reading the stream as a caller