Why you should build agents without tools first

Tools add latency, schema risk, and surface area to debug. Before you reach for them, see if the task is actually a sequence of pure-language transformations: classify, extract, summarize, rewrite, score.

If yes, the agent is a loop of LLM calls — no tools needed. You'll ship faster and the failure modes are easier to reason about.

Reach for tools when (a) you genuinely need to act on the outside world, (b) the data lives outside the prompt and is too large to embed, or (c) you need verifiable side effects with audit trails.