Workflow (in AI)

In Anthropic's "Building effective agents" framing, a workflow is an AI system where the steps and their order are decided by the developer at design time. The LLM fills in each step, but it doesn't choose what comes next. Step 1 always runs, then step 2, then a conditional, then step 3. Like a flowchart with LLMs in the boxes. Workflows are rails; agents wander. That distinction matters. A workflow is more predictable, easier to evaluate (you can test each step), cheaper (no exploration loops), and more debuggable (you know exactly where it failed). An agent is more flexible — it handles cases you didn't anticipate — but harder to control. Most production AI in 2026 is workflows, even when it's marketed as "agentic." Common workflow patterns: prompt chaining (output of step 1 → input of step 2), routing (classify the input, then call a different prompt depending on class), parallelization (fan out to multiple LLM calls, merge results), evaluator-optimizer (one model generates, another critiques, loop until acceptable). These cover 80% of useful AI applications. The agentic loop only earns its keep when the task is genuinely open-ended. Practically: start with a workflow. Sketch the steps, hardcode the order, see if it works. Move to agentic only when you genuinely don't know the steps in advance — when the model needs to decide whether to search the web first or check the database first based on the question. Most "agents" you read about are workflows with one LLM call wearing a hat, and that's fine — workflows that ship beat agents that demo.