Agent Workflow Pattern (Trigger → Input → Action → Output)

What It Is

A factory agent task is specified as four explicit parts so it can run unattended and be gated:

  1. Trigger — the event that starts it (e.g. a draft passes the voice/fact gates).
  2. Input — the defined data it consumes (e.g. the draft + 5–10 personas built from real buyer profiles).
  3. Action — what it does (e.g. each persona produces a textual reaction; SSR maps reactions to Likert distributions).
  4. Output — a structured result that proceeds and is logged (e.g. a comparative ranking of variants + qualitative reactions; the winner publishes; the run is logged as an experiment row).

Worked Example: the SSR Pre-Publication Test

The synthetic-consumer-panels gate is a clean instance: trigger = draft passed gates; input = draft + personas; action = elicit text → SSR; output = ranked variants + reasons, logged for the calibration-loop. The same skeleton describes content, outbound, analytics, and community agents — only the trigger/input/action/output contents change.

How It Applies to Marketing Factory

This is the unit of composition for the whole factory. Writing every agent in trigger→input→action→output form makes tasks gateable, loggable, and swappable, and it makes the human/agent split explicit — execution lives in the Action, while strategy and judgment stay at the gates (see agent-ownership-boundary). Every Output should be a logged, structured row so the factory accumulates its own evaluation data.

Referenced from: ssr-synthetic-panels