'Agent' is an overused word. At its core, an agent is an LLM in a loop with tools: it decides what to do, takes an action via a tool, observes the result, and repeats until the task is done. Here's how that actually works.
The Agentic Loop
- The model receives a goal and a set of tools it can call.
- It decides on an action and emits a tool call (e.g. search, run code, query a DB).
- Your code executes the tool and returns the result to the model.
- The model reasons about the result and either calls another tool or finishes.
Tools Are the Agent's Hands
An agent is only as capable as its tools. Give it a search, a read_file, a run_query — each with a clear description of *when* to use it. The model picks tools based on those descriptions, so write them well.
When NOT to Build an Agent
Agents add cost, latency, and unpredictability. If a task is a single well-defined step (classify, summarize, extract), use one LLM call. If it's a fixed multi-step pipeline, orchestrate it in code. Reach for an agent only when the path genuinely can't be specified in advance.
Guardrails Are Mandatory
An agent that can run code or send emails can do real damage. Add iteration limits, gate destructive actions behind confirmation, and log every tool call.
