Drug-Safety Agent on Databricks · Business flow · Why agentic
Why agentic? An honest answer
The word agent is easy to over-claim. For one named drug and side effect, most of this work is the same steps every time, and a script would do it. A model earns its place only where the path is not fixed. Here is the honest split.
| Step | Needs a model? | Verdict |
|---|---|---|
| Fetch counts, compute PRR, check the label | No. The same steps for every pair. | A plain script does it better. In this build the model still calls these tools, which is a choice we could revisit. |
| Pick which pairs deserve a draft | No. A rule: a signal not already on the label. | Plain code, built. |
| Refuse causation, advice and filing | No. Rules are safer than manners. | Rules, built. |
| Check every number against the data | No. Comparison. | Plain code, built. |
| Understand an open question such as what should I look at for amiodarone? | Yes. The path is not fixed. | This is where an agent is justified. Not built or tested yet. |
| Write the assessment, uncertainty and recommendation in plain words | Yes, with the numbers fixed by code. | A model inside guard rails. Built. |
| Read label or case narratives for context | Yes. | Future. |
What this means. Today the agent is a deterministic core with a model at the edge that writes the words. For a single named pair a template could do about nine tenths of it. The model becomes necessary for open questions and for text, and the open-question case is the one not yet built.
The position to take. Deterministic first; agentic only where it is a must. This build follows that: rules route, code computes, checks verify, and the model is used where wording or an open question needs it. That is a more defensible claim than "it is an agent".
Why it is still worth building as an agent. The same harness (limits, checks, approval, audit) is what makes the open-question case safe to add later. Building the safe frame first is the point of the exercise.