Marketing Agent Orchestration — The Conductor Layer

For factory operators. The KB has the instruments — content, outbound, analytics, community, paid, and SEO agents, each with its own agent-workflow-pattern. What it lacks is the score: the layer that sequences them, passes state between them, routes work, handles failure, and decides when the job is done. This document is that conductor — the difference between "six agents" and "a factory."

The thesis: orchestration is mostly plumbing and judgment, not intelligence — and the first rule is to use the least of it that works. Most marketing work is a workflow (predefined path), not an open-ended agent; reaching for autonomy first is the most common and most expensive mistake.

1. Workflows vs. Agents (the load-bearing distinction)

(Anthropic, Building Effective Agents.)

  • Workflows — "systems where LLMs and tools are orchestrated through predefined code paths." Predictable, debuggable, cheap.
  • Agents — "systems where LLMs dynamically direct their own processes and tool usage." Flexible, but unpredictable and costly.

The governing principle: "add complexity only when it demonstrably improves outcomes." Optimize a single LLM call first; escalate to a workflow only when the task needs it; escalate to a true agent only for open-ended problems where the steps can't be predicted in advance. For a marketing factory, the overwhelming majority of work is a workflow — the content calendar, the nurture sequence, the weekly report — and should be coded as one.

2. The Five Canonical Patterns (the ladder)

Pattern What it does Marketing use
Prompt chaining Sequential steps; each call processes the prior output, with gates between Draft copy → translate; outline → write → fact-check
Routing Classify input, send to a specialized handler Inbound query → general / refund / technical agent; lead → SMB vs enterprise track
Parallelization Sectioning (independent subtasks) or voting (run N times) Parallel guardrail + compliance + brand-voice screens; multi-aspect content scoring
Orchestrator-workers A central LLM decomposes, delegates to workers, synthesizes A campaign brief fanned out to content + design + landing-page agents, then assembled
Evaluator-optimizer One agent generates, another critiques in a loop Copy refined against a brand-voice/quality rubric until it passes

The orchestrator-workers pattern is the most widely deployed in production: the orchestrator plans, routes, and supervises but does not do the work itself — it picks the next agent, enforces handoff rules, and decides termination. Use it when subtasks can't be enumerated up front (e.g. a research-and-produce campaign); use plain chaining/routing when they can.

3. Shared State & Memory (what makes hand-offs work)

Agents only collaborate if context survives between them. The orchestration substrate:

  • Shared state / scratchpad / blackboard — a workspace object that persists across hand-offs, holding the campaign brief, intermediate artifacts, and decisions. Without it, each agent starts blind.
  • Short-term memory — the current session/run's working data.
  • Long-term memory — knowledge persisted across runs (brand voice, past-campaign outcomes, ICP). For a factory, the KB itself is long-term memory; experiment logs feed it.
  • Handoff protocol — explicit rules for what one agent passes to the next and when control transfers. Ambiguous hand-offs are the #1 source of multi-agent failure.

Topologies: hierarchical / orchestrator-worker (one supervisor, clear control) vs swarm / blackboard (peers share memory, hand off peer-to-peer). Hierarchical is easier to reason about and the right default for a marketing factory; swarm suits loosely-coupled, exploratory work.

4. Orchestration Meets Governance

Orchestration is where the factory's safety rails are enforced, not bypassed. Every hand-off is a place to insert a human-review-gate; every worker output passes the output-validation/guardrail step before it propagates; the orchestrator is responsible for detecting and escalating agentic-failure-modes (a worker timing out, drifting, or returning low-confidence). Parallelized guardrail screens (compliance, brand-voice, prompt-injection) are a canonical use of the parallelization pattern.

5. Factory Integration

The conductor is the spine of "a marketing factory that agents can run": it sequences the per-channel agents (each an agent-workflow-pattern) into campaigns, holds the shared brief/state, routes work by type and difficulty, runs evaluator-optimizer loops for quality, and decides termination — while keeping humans at the gates. The design discipline is restraint: model each motion as the simplest pattern that works (usually chaining or routing), reserve orchestrator-workers for genuinely unpredictable jobs, and reserve full autonomy for almost nothing. This is the assembly layer above content-machine, outbound, and analytics agents that turns the marketing-factory from a metaphor into a runnable system.

Provenance Note

  • Primary-sourced, fetched & verified: the workflow-vs-agent definitions, the five patterns (prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer), and the "add complexity only when it demonstrably improves outcomes" principle (Anthropic, Building Effective Agents, fetched).
  • Secondary — search summary, not independently fetched: the shared-state/blackboard, handoff-protocol, and hierarchical-vs-swarm topology framing (orchestration practitioner blogs; Microsoft Azure agent design-patterns). Mechanism is well-established; specifics not promoted to verified.

Sources

Concepts

Extracted from this source: agent-orchestration · workflow-vs-agent · shared-agent-memory

Related concepts: agent-workflow-pattern · human-review-gate · agentic-failure-modes · marketing-factory · content-machine