Workflow

Ask an agent to do something once and you get a single response. Give it a workflow, and you're describing a repeatable sequence: do this, then that, then check something, then produce a result. The term predates AI agents by decades — it's the same word used for business processes and CI pipelines — but in an agent context it usually means the sequence of steps an agent (or a system built around one) follows to get from a request to an outcome.

A workflow doesn't have to be fully automated. It's common for a workflow to include a step where a human reviews or approves something before the next step runs — the agent handles the mechanical parts, a person handles judgment calls.

Workflows can be described informally in prose (a playbook), encoded as an explicit sequence a framework executes (a pipeline), or, when steps can run in parallel and later need to converge, modeled as a dag-workflow. “Workflow” is the general umbrella term that covers all of these.

How it works

Concretely, a workflow starts from a trigger — a new support ticket, a schedule, a manual command — and then runs through its defined steps in order. Each step can be an agent doing work, a script running, or a human approving something. The workflow ends when the final step produces its result, which might be a merged PR, a sent report, or an updated record.

Example

A content-review workflow for a blog might run: a new draft is submitted (trigger) → an agent checks it against a style guide (step 1) → an editor approves or requests changes (step 2) → the post is published (result).

How it differs

Common misconceptions

FAQ