Agents with real tool use

An agent is a control loop around a model. The model proposes an action, a tool runs it, the result comes back, and the model decides what to do next, until the task is done or a limit trips. That’s the whole trick. Everything else is engineering around it.

The tools are where the real work happens: query the warehouse, call an internal API, write the file. We build most of them as MCP servers now, one protocol between models and systems instead of a custom integration per pair, which also means the same tools serve Claude today and whatever replaces it. The model chooses the action. The server decides what’s allowed, and nothing runs without permission.

The guardrails are plain engineering: step caps, cost caps, argument validation before anything executes, a log of every action. And evaluation on real tasks, rerun every time the loop changes, because an agent that was reliable in March isn’t automatically reliable in June.