In short
Workflow patterns are familiar ways to arrange steps: a chain, a route, parallel tasks, a review loop, or a coordinator. Choose the smallest pattern that makes the job clear, and give every loop a stopping point.
Why it matters
Two workflows can use the same assistant and behave very differently. One might pass a draft through defined checks. Another may let the assistant decide what to investigate next. Naming the pattern helps you see where information can be lost and where a person should intervene. It also makes a process easier to explain to someone who will operate it.
How it works
A chain passes an output into a later step, such as extract, check, then format. Routing sends different kinds of input down different paths. Parallel work handles independent pieces at the same time and then joins them. A review loop checks an output against criteria and may request a revision. Orchestration assigns work and brings the parts together when the subtasks cannot all be named in advance. Anthropic’s 2024 guide describes these patterns as practical building blocks, with tradeoffs in cost, speed, and complexity.
Pick based on the actual variation in the job. If the input always has the same fields and the output is a fixed table, a chain may suffice. If it can be a price inquiry or a technical question, route by category and include an uncertain category. Parallelize only when the pieces can be checked independently; otherwise the joining step may hide an unresolved conflict. A review loop needs criteria that can identify a problem and a maximum number of attempts. A coordinator needs explicit roles and a way to preserve the source of each claim.
NIST’s 2023 AI Risk Management Framework treats measurement and management as continuing activities. Applied to patterns, this means checking the path as well as the final text. Track which branch was chosen, whether the reviewer corrected it, and whether exceptions reached a person. Do not accept a diagram as evidence that the process works when its handoffs are opaque.
Example
Illustrative example: this scenario is invented.
Sam, 19, helps plan a neighborhood game afternoon. They collect public suggestions for activities. A routing step separates indoor and outdoor ideas, and a chain turns each group into a short materials list. One suggestion lacks a location, so it goes to an undecided list for Sam to sort. They do not let the assistant reserve space or invite people. The simple route and chain handle the known categories; an agent choosing many new steps would add little here.
What it doesn’t mean
A review loop isn’t a guarantee of improvement; it can repeat the same error. Parallel answers aren’t independent if they rely on one mistaken source. Routing shouldn’t force an ambiguous item into a confident category. Every added pattern should solve a real problem that a simpler flow could not handle.
Try it
Take six imaginary public-event suggestions on paper. Ask any assistant to describe a chain and a route that would sort them, including a path for unclear entries. Draw where a person would see the original suggestion. Decide which pattern keeps the task understandable with the fewest steps.
Keep going
- Designing an AI workflow
- When more than one agent helps
- The human in the loop needs a real job.
- Map a workflow before connecting apps
Sources
- Building effective agents, Anthropic, 2024. Checked October 13, 2026.
- AI RMF Core, NIST, 2023. Checked October 13, 2026.