Lesson 1 of 7
What is an agent (vs a workflow)
7 min read
Everyone ships "agents" now. Most of them are workflows in a trench coat. What's the actual difference — and why does it decide how you build?
Workflow: you write the steps
A workflow is a fixed path you design: the LLM fills in each step, but you wrote the control flow. Prompt chaining, routing, parallelization — the shape is decided before the task runs. Predictable, cheap to debug, easy to trust. Anthropic's Building Effective Agents catalogs five such shapes.
A workflow is orchestration you control: the model does the sub-tasks, but the steps are yours. Most production "AI features" are workflows — and that's a strength, not a shortfall.
Agent: it writes the steps
An agent is different: you give it a goal and tools, and it decides the steps — how many, in what order, when it's done. The control flow is dynamic, discovered at runtime. That's more capable and far less predictable: it can loop, cost more, and surprise you.
The line is who owns the control flow. Workflow: you do. Agent: the model does. Start with a workflow; reach for an agent only when the task genuinely needs runtime decisions.
"Agentic" is a spectrum, not a binary. A workflow with one dynamic routing step is more agentic than a straight chain, less than a full loop. Pick the least agency that solves the task.
The shape of it
- —Workflow — you design the steps; the LLM fills each one in.
- —Agent — you give a goal plus tools; it decides the steps at runtime.
- —The question is who owns the control flow, not how smart the model is.
- —Default to a workflow; add agency only where the task demands it.
Your feature always runs the same three steps in the same order, every time. What have you built?
Continue in the app
Take the whole AI Agents in Production course — tracked
Get your personalized path, progress and streaks in the app — this lesson and every next one, in order.