Back to blog
AI•June 10, 2026•4 min read

Agents and tools: when to let the model act

The moment you hand a model tools-let it call an API, run code, write to a database-it stops being a thing that answers and becomes a thing that acts. That unlocks real agents, but it also means a wrong decision now has consequences beyond a bad sentence. The interesting design work isn't "can it use tools," it's deciding which actions it gets and what guards sit around them.

Start with the simplest thing that works

Not everything needs to be an agent. A single well-prompted call handles most classification, extraction, and Q&A. A fixed pipeline-where your code orchestrates the steps and the model fills in the judgment-covers most of the rest. Reach for an open-ended agent only when the task is genuinely multi-step and you can't specify the path in advance. Agents cost more in latency, tokens, and unpredictability; pay that only when the task earns it.

The power of tool use is also its risk. Promote an action to a tool so you can gate it-not just so the model can reach it.

Gate the actions that bite

The reason to make something a dedicated tool, rather than letting the model run free, is control: a typed send_email or refund_payment tool is something your code can validate, log, and-when it matters-pause for human approval. Reversible, low-stakes actions can run automatically; irreversible or high-stakes ones should stop at a gate. That's the same human-in-the-loop principle that makes AI features trustworthy, applied to what the model is allowed to do.

A checklist before you let a model act:

  • Could a single call or a fixed pipeline do this? If yes, don't build an agent.
  • Is the action reversible? Reversible can be automatic; irreversible needs a gate.
  • Can you validate the inputs before executing? Make it a typed tool so you can.
  • Will you see what it did? Log every tool call-an agent you can't audit is one you can't trust.