Workflows and agents

📘What it really is

• Workflow: your code fixes the steps, and the model does specific parts. For example: classify the enquiry (model) → look up the customer (code) → draft a reply (model) → a person approves (human) → send (code).

• predictable, easy to test, cheaper; best for most business processes.

• Agent: the model decides which tools to use, in what order, and when it's done, within limits you set.

• flexible for open-ended tasks ("find out why this customer's orders dropped and suggest actions");

• less predictable, harder to test, can take more steps and cost more; mistakes can compound.

• Start with the simplest thing that works: a single model call, then a workflow, and only then an agent, if the task really needs open-ended exploration and mistakes can be caught and recovered (tests, review, undo).

• Common workflow patterns:

• chaining: one step's output feeds the next;
• routing: classify the request, then send it to the right specialised prompt or tool;
• parallel: do several independent checks at once and combine them;

• evaluator: one step drafts, another checks against criteria.

• Agents need guardrails: step limits, cost limits, timeouts, a small set of tools, approval for actions (Lesson 5) and full logs of every step.