Five agentic workflow patterns, drawn at one scale
By Christophe Paka · Facts checked on 23 October 2026
A vendor shows a diagram full of boxes and arrows, with the word agent in the title. Is it one? Most of what ships under that word is one of five agentic workflow patterns from Anthropic's guide, Building effective agents.
Miss the pattern and you pay agent costs for a fixed chain of steps. A workflow is a set of steps you wrote in code; the AI model only does the work inside each step. Learn the five shapes and you can name a system before you read its brand. The figure draws all five at one scale.
How to read this figure
Panel 1, prompt chaining. Model call 1, then a diamond called the gate, then model call 2, then the answer. A model call is one request to the AI model. The gate is a check your code runs: on fail, the chain stops. Same steps, every run.
Panel 2, routing. A router classifies the input first, then sends it down path A, B or C. One input, one chosen path. Use it when inputs fall into groups that are better handled apart.
Panel 3, parallelization. Three calls run at once, and an aggregator combines their answers. The pattern has two named variants. Sectioning runs three different subtasks, which buys speed. Voting runs the same task three times, which buys confidence.
Panel 4, orchestrator workers. An orchestrator hands work to three workers along dashed arrows, and a synthesiser combines their results. Dashed means how many workers, and what each does, is decided live, during the run. That makes it the pattern closest to an agent.
Panel 5, evaluator optimizer. A generator drafts, an evaluator critiques, and a feedback arrow sends the draft back until it meets the criteria. Then it goes out as accepted. You wrote that feedback arrow, in code.
Panel 6, the shaded box. The orange panel says all five are workflows. You wrote every arrow, and the model fills in the boxes. An agent has no fixed arrows: the model decides what happens next while the run is happening. Compare the same job as a workflow and as an agent.
Key takeaways
- Name the pattern before you call a system an agent.
- Point at each arrow of a diagram and ask who drew it, you or the model.
- Choose voting when you need confidence in one answer, and sectioning when you need speed.
Worked example: five jobs, five patterns
Anthropic's guide gives one example job for each pattern. Match each job to its panel in the figure.
- Write marketing copy, then translate it: prompt chaining, because the task splits into fixed subtasks.
- Send easy questions to a small model and hard ones to a larger model: routing.
- Review code for security flaws with several prompts: parallelization, the voting variant.
- Make a code change that spans many files: orchestrator workers, because you cannot predict the subtasks in advance.
- Translate a literary text with a critic in the loop: evaluator optimizer, because clear criteria exist and a second pass helps.
Result: All five jobs run on arrows you wrote. Only the change across many files lets the model pick its subtasks, and even there you wrote the frame: the orchestrator, the workers and the final combine step.
Questions readers ask
At what point does a workflow become an AI agent?
When any arrow appears only once the run starts. If every arrow is yours, it is a workflow, whatever the project calls itself. Orchestrator workers comes closest: the model picks the subtasks, but you still wrote the orchestrator, the workers and the step that combines them.
Is anything with an LLM in it an agent?
No. An LLM, or large language model, is the kind of AI behind ChatGPT and Claude. A prompt chain calls one at each step, but your code sets the order. OpenAI's guide says simple chatbots and sentiment classifiers are not agents, because the model does not control the order of the steps.
Sources: Building effective agents (Anthropic); A practical guide to building agents (OpenAI, PDF).
This figure is from AI Agents Made Crystal Clear, chapter 2.
The chapter builds the idea step by step, with recipes you can follow, a Try it to check yourself, and every source.
Get the book
