Vol. I · First Principles · I.2
Workflows versus Agents
Anthropic draws a bright line: workflows orchestrate LLMs through predefined code paths; agents let the model direct its own process. Most production systems that work are workflows with a small agentic island.
7 min read
Doctrine
- Start with the simplest thing that works: one prompt, then a workflow, then an agent.
- A workflow is not a lesser agent. It is the correct architecture when the graph is knowable.
- Frameworks that collapse the distinction force you to reverse-engineer them at 80% quality.
The Anthropic catalog
Building Effective Agents (Anthropic, 2024, still the field's north star) names five workflows — prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer — and one autonomous agent pattern. The agent is last because it is the most expensive and the most failure-prone.
The test is simple. Can you name the steps? Use a workflow. Can you not name them because the environment talks back (a repo, a browser, a ticket queue)? Use an agent, with a budget and a sandbox.
What production actually looks like
Horthy's survey of founders: the products billed as agents are mostly deterministic code with LLM steps at the points of ambiguity. n8n and Dify industrialize that shape. LangGraph lets you mix — deterministic edges with an agent node where the path fans out.
CrewAI learned this the hard way and added Flows: the role-play is for the fuzzy middle; the flow is for the parts you must not get wrong.
Anti-patterns
- Reaching for a multi-agent crew because the demo looked collaborative.
- Encoding a five-step invoice pipeline as a free-running ReAct loop.
- Assuming more LLM calls equal more intelligence.