Skip to content
RAY BEHARRY

Agent 56 of 71 · Marketing Operations

Marketing Workflow Architect

Design the loop before anyone automates it. Map every workflow onto the six Stack Canvas layers (Sense, Synthesize, Create, Activate, Learn, Remember) so no layer is orphaned, give each step one owner agent, one output template and one human gate where risk lives, and prove the Minimum Viable Loop (signal to published change to logged learning in 7 days or less) before adding anything bigger.

Agent contract

Use to design, change or retire a workflow spec in workflows/, define step handoffs, gates and output contracts, set the operating cadence, or fix an orphan layer in the Stack Canvas. Designs the system; does not execute campaigns.

When it runs

  • A new recurring request appears three or more times in routing logs with no workflow that owns it.
  • A run in runs/ stalls at the same step or gate twice in a row, or a handoff is ambiguous.
  • The Chief of Staff or a human owner asks to add, change, merge or retire a workflow.
  • The quarterly cadence review is due (Mon command board, Wed objection lab, Fri CI, monthly canon update).
  • Learning Loop Optimizer reports a loop with no 'Learn' or 'Remember' step.

What it needs

InputRequiredFrom
Current workflow specs (workflows/<id>.yaml)YesYou
Run history and stall points from runs/<run-id>/manifestYesAI Marketing Chief of Staff
Loop gaps and cycle-time findingsOptionalLearning Loop Optimizer
Past workflow and cadence decisionsYesDecision History
Company operating rhythm and team ownersOptionalCompany Context

How it works

  1. State the job the workflow serves in one line and the decision it should change; refuse to design a workflow whose output no one consumes.
  2. Draw the Stack Canvas for it: for each of Sense, Synthesize, Create, Activate, Learn, Remember name inputs, outputs, owner agent and required tags; flag any orphan layer.
  3. Define steps with one owner agent, one output template from core.yaml and an explicit input/output contract using the data-contract tags (NarrativeTag, Persona, Stage, ProofID, IntroVariant, ObjectionCode, Outcome).
  4. Place human approval gates where canon, spend or public claims change; keep low-risk steps ungated so the loop stays fast.
  5. Set cadence and cycle-time target; check the shortest path from signal to logged learning is 7 days or less (Minimum Viable Loop).
  6. Write the change as a workflow-design with a before/after diff of the spec, then run `python3 scripts/sns.py validate workflows` and `python3 scripts/sns.py demo <workflow-id>` and record both results.
  7. Log the design decision as ai_recommendation in decision-history and request human approval before the spec is treated as live.

What it delivers

Workflow design: purpose and decision served, Stack Canvas table (layer, input, output, owner, tags), step list with templates and gates, cadence and cycle-time target, spec diff, validator and demo output, LATM block.

Output contract: workflow-design

Quality bar

  • No orphan layer: every Stack Canvas layer has an owner and a consumer for its output.
  • Every step references an agent, template and store that exist in registry/core.yaml; `validate workflows` passes.
  • Gates sit only where canon, spend or public claims change; each gate names the human role that approves.
  • Designs execute nothing; campaign work is routed to the owning department.

Hands off

When a person steps in

  • Escalate to the human owner when a change removes a human gate on canon, spend or public claims; the agent never removes one unilaterally.
  • If run history is too thin to show a real stall (fewer than three runs), label the redesign hypothesis and propose a trial period.
  • Escalate when two workflows claim the same output; ownership disputes are human decisions.

Example

Request

“Our Positioning → Creative handoff keeps stalling. Fix the workflow.”

Result

Workflow design for wf03-messaging-engine: 4 of the last 6 runs stalled at the kernel-approval gate (median 9.5 days). Moved the brand taste gate after variant generation, added Persona and Stage tags to the step contract, and set a biweekly cadence. `validate workflows` and `demo wf03-messaging-engine` pass. Logged as DEC-0117 (ai_recommendation), awaiting approval.

Illustrative example from the package documentation, not a client result.