When an agent is the right design
An agent is useful when the path to an answer varies from case to case: researching an account, triaging a messy inbox, reconciling records across systems. If the steps are fixed, a normal workflow or n8n flow is cheaper and easier to audit. We often start by drawing the task and labelling which steps truly need judgment. Mapping the task first also tells us how much autonomy the process can tolerate.
That exercise usually yields a hybrid. Deterministic code handles routing, validation and writes, while the model handles interpretation and choices inside a bounded space. Smaller, narrower agents that hand off to each other tend to be easier to test than one large agent with every tool attached. Each agent gets a short written contract describing its inputs, outputs, tools and the situations where it must stop and ask for help.
