Agentic workflow

The distinction most practitioners now use comes from Anthropic's "Building effective agents" (December 2024): workflows are systems in which LLMs and tools are orchestrated through predefined code paths, whereas agents let the model dynamically direct its own process and tool use. In a workflow the control-flow graph is authored by the developer and is known before execution; the model fills in nodes. The term "agentic workflow" is used loosely in industry, sometimes for anything involving an LLM loop, so it pays to ask who owns the next-step decision: code or model.

The same article names five recurring patterns. Prompt chaining decomposes a task into sequential calls, with programmatic gates between steps (for example, rejecting a draft that fails a length or schema check). Routing classifies the input and dispatches it to a specialised prompt or cheaper model. Parallelisation runs calls concurrently, either by sectioning a task into independent subtasks or by voting, running the same task several times and aggregating, which trades cost for reliability. Orchestrator-workers lets a central LLM decide at run time how to split the work, and evaluator-optimiser loops a generator against a critic until acceptance criteria are met. The last two already hand some control to the model and sit on the boundary with agents.

Engineering a workflow looks like ordinary distributed-systems work. Each step should exchange structured output validated against a schema, so failures are caught at the edge rather than propagating as free text. Steps need timeouts, retries with backoff, idempotency keys for side-effecting tool calls, and checkpointing so a long run can resume; durable-execution engines and graph frameworks such as LangGraph exist largely for this. Because each node has a narrow contract, it can be evaluated in isolation with its own test set, which is the main practical advantage over free-running agents, whose trajectories vary from run to run.

Workflows also make control placement explicit. A human-in-the-loop approval can be inserted at the one step that sends email or moves money; untrusted content can be confined to steps that have no dangerous tools; and cost and latency are bounded because the number of model calls is fixed or capped. The failure mode is rigidity: inputs the designer did not anticipate fall through the routing logic. A common evolution is to start with a chain, add routing as edge cases appear, and grant agentic autonomy only to the sub-task that genuinely needs open-ended exploration.