When to Build an Agent vs When a Single Prompt Is Enough

A single prompt-and-response call is enough when the task is one bounded step: answer a question, draft something, classify something, transform some input into some output. An agent — a model that plans, calls tools, checks its own results, and decides what to do next in a loop — earns its complexity when the task genuinely can't be done in one step: - The number of steps needed isn't known in advance ("research this topic" might take one search or ten, depending on what it finds). - Later steps depend on the results of earlier ones in a way you can't predict and hard-code (branching investigation, iterative debugging, multi-tool workflows). - The task needs the model to notice its own mistake and try a different approach, not just produce one answer and stop. Signs you've reached for an agent too early: you can already write down the fixed sequence of steps by hand (that's a workflow/PromptBook, not an agent), the task completes in one or two tool calls every time (that's just function calling), or you're adding an agent loop to work around a prompt that could be fixed directly. Agents add real cost — more tokens, more latency, more ways to fail silently by looping or going off-task — so the honest first question is always "could a simpler thing already do this," not "how do I build the agent version."