Workflow vs. Agent (Start Simple)¶
What It Is¶
A distinction from Anthropic's Building Effective Agents:
- Workflow — "systems where LLMs and tools are orchestrated through predefined code paths." Predictable, debuggable, cheap.
- Agent — "systems where LLMs dynamically direct their own processes and tool usage." Flexible, but unpredictable and expensive.
The governing rule: add complexity only when it demonstrably improves outcomes. Optimize a single LLM call first; escalate to a coded workflow only when the task needs multiple steps; reserve true agentic autonomy for open-ended problems where the steps genuinely can't be predicted.
The Complexity Ladder¶
single call → prompt chaining → routing → parallelization → orchestrator-workers → autonomous agent. Each rung adds capability and unpredictability and cost. You climb only when the rung below demonstrably fails the task.
How It Applies to Marketing Factory¶
This is the most important discipline in agent-orchestration, because the default failure mode is over-engineering: reaching for an autonomous agent where a three-step workflow would be more reliable and far cheaper. The overwhelming majority of factory work — content calendars, nurture sequences, weekly reports — is a workflow and should be coded as one, with humans at the human-review-gate. Choosing the right rung is also a safety decision: less autonomy means fewer agentic-failure-modes and a smaller blast radius. Each rung is built from agent-workflow-pattern units.
Related Concepts¶
- agent-orchestration — workflow-vs-agent governs how much orchestration to apply
- agent-workflow-pattern — the unit each rung composes
- agentic-failure-modes — more autonomy widens the failure surface
- human-review-gate — workflows keep humans at fixed decision points
Referenced from: marketing-agent-orchestration