Agentic workflow; agent orchestration
An AI assembly line that routes a job through model calls, tools, branches, and review steps instead of asking one model once.
See it
What it is
An agentic workflow, also called agent orchestration, coordinates several model calls, tools, branches, and checkpoints into one job. One step may retrieve documents, another draft, a critic score the draft, and a final step revise it. Some routes are fixed like a directed acyclic graph; others let a model choose the next node at runtime.
Reach for it when a task has distinct stages with different context, tools, or quality gates. Put predictable work in explicit workflow nodes and reserve agent decisions for the branches you truly cannot know in advance. Anthropic's 'Building Effective Agents' post drew the line worth keeping: a workflow follows a path you wrote, an agent picks its own. Named inputs and outputs between nodes make the run testable without replaying the entire system.
Gotcha: orchestration does not make unreliable steps reliable. Retries can send the same email twice, parallel branches can disagree, and a cycle can quietly burn the budget. Persist state, make side effects idempotent, cap every loop, and trace each node with its inputs, outputs, model, tools, cost, and stop reason.
Ask AI for it
Build this task as an explicit directed acyclic graph with nodes named intake, retrieve, draft, parallel_fact_check, critic, revise, approval, and publish. Define every node's input and output with JSON Schema, run independent fact checks with Promise.allSettled, and route failures to a repair node instead of restarting the graph. Persist node state after every transition, use exponential backoff only for transient errors, attach an idempotency key to every side effect, cap revision at two passes, require human approval before publish, and emit an OpenTelemetry span containing model, token count, latency, and outcome for every node.