Drug-Safety Agent on Databricks · Architecture
The architecture in one picture
Every agent that does real work needs the same sixteen jobs done around the model: assembling what it reads, calling it, running tools, looping, remembering, checking, asking a person, and more. We call this surrounding code the harness. This page shows the sixteen jobs and how much of each this agent has.
Most of it is borrowed, on purpose. The loop, the limits, the tool contracts and the model connection come from a shared library, agent_harness, that every agent in this family uses. What this agent writes is only what is particular to drug safety: the rules, the three tools, the statistics, the output shape and the grounding check. Read the table under Sixteen layers, a table for each layer.
Why show the gaps. A tour that only shows what works teaches little. One layer is not built at all (observability). Two more, state and human approval, exist only as local stand-ins: a SQLite store and a web page without a login. The status page puts the gaps in order.