Plans First: Why AI Should Think Before It Acts

There are two ways to connect an AI agent to your systems. The first is direct: give the model tool definitions, let it decide which tools to call, and execute each call as it's generated. The second is indirect: let the model produce a full plan first, validate that plan, and then execute it step by step.

Most teams default to the first approach because it's simpler to set up. But the second approach — plans first — is significantly more reliable, safer, and easier to debug.

When an AI agent calls tools directly as it reasons, each step depends on the last. If the model misreads the output of an API call in step two, steps three, four, and five are all built on that mistake. By the time you notice something went wrong, the agent may have made a dozen downstream calls based on a faulty assumption.

Direct execution also makes it hard to pause. Once the model starts calling tools, stopping it mid-stream means interrupting an in-progress operation that may have already made irreversible changes.

A plan is a structured, inspectable description of what the agent intends to do — before it does anything. Instead of "call this endpoint, then based on the result, call this other endpoint," a plan is a complete sequence: here are all the steps, here are the dependencies between them, here are the inputs each step expects.

This is something language models are naturally good at. Describing intent in natural language and structured steps is much closer to how models are trained than generating perfectly valid API calls with correct schemas and authentication.

Once you have a plan as a structured object, you can do things you can't do with direct execution. You can validate every step against your registered task schemas before anything runs. You can check for policy violations — does this plan try to call a restricted endpoint? Does it attempt to access data the requesting user isn't allowed to see?

Most importantly, you can show the plan to a human and ask for approval. A plan is readable. A live stream of API calls is not.

When a plan works, you can save it. The next time a similar task comes up, the agent can start from a proven plan rather than generating a new one from scratch. This dramatically reduces hallucinated steps and makes agent behavior more consistent over time.

Think of it like the difference between asking a new employee to figure out a process on their own every time versus giving them a runbook. The runbook can be improved, but it prevents reinventing the wheel — and catches mistakes before they become incidents.