Skip to main content
A draft is generated, the workflow pauses to ask the operator to approve it, and only then does the publisher run. This is the canonical human-in-the-loop pattern: an autonomous pipeline with one explicit gate where a human decides whether to continue.

​
What it demonstrates

  • The human node type
  • Resolving {{ draft.output }} inside the prompt shown to the operator
  • A when: edge that only routes to publish if the human answered "yes"
  • timeout + on_timeout: use_default so an unattended workflow doesn’t hang forever

​
Run it

sirenspec run docs/cookbook/content-approval/workflow.yaml \
  --input "A weekend launch announcement for our new YAML-first agent SDK."
The CLI will print the draft, then block on stdin. Type yes to publish or no to reject.

​
Workflow

docs/cookbook/content-approval/workflow.yaml
version: "0.1"

agents:
  drafter:
    model: "openai:gpt-4o-mini"
    system: |
      You are a marketing copywriter.  Draft a short, punchy social media post
      about the topic the user provides.  Return only the post text — no preamble.

  publisher:
    model: "anthropic:claude-haiku-4-5-20251001"
    system: |
      You are the publishing assistant.  Format the approved draft into a final
      post with a one-line caption and 3 hashtags.

nodes:
  draft:
    agent: drafter
    writes: working.draft

  approve:
    type: human
    prompt: |
      Draft to review:
      ---
      {{ draft.output }}
      ---
      Reply 'yes' to publish, 'no' to reject, or any text to edit (free-form).
    writes: working.approval
    timeout: 600          # 10 minutes
    on_timeout: use_default
    default_output: "no"

  publish:
    agent: publisher
    writes: output.post

edges:
  - from: draft
    to: approve
  - from: approve
    to: publish
    when: 'working.approval == "yes"'
human nodes consume no LLM tokens. They block on stdin (or any caller-supplied input function), and the operator’s response is written to the workflow context just like an agent’s output — so downstream when: conditions, prompt interpolation, and writes paths all work identically.

​
Timeout semantics

on_timeoutBehaviour when the timeout fires
abortThe workflow fails with a HumanInputError.
skipThe empty string is written and execution continues.
use_defaultdefault_output is written and execution continues.
Setting on_timeout: use_default with default_output: "no" (as above) makes the workflow fail-closed: an unattended pipeline never auto-publishes.

​
Graph

​
Next steps

Budget Guarded

Cap a workflow’s total spend with the budget: block.

Conditional Pipeline

Route to different handlers based on an LLM classification.