Most AI demos put the agent at the centre and hope.
For regulated work that's backwards.
The core has to be deterministic. Same inputs, same path, every time. Reconstructable six months later when someone asks why the system did what it did.
An audit trail is only worth something if the thing it describes was going to happen the same way twice.
Agents belong at the edges. Reading the messy document. Drafting the summary. Handling what doesn't fit the schema. Every one of those calls recorded with its provenance: what went in, what came back, which version decided.
Rules at the centre. Judgement at the edges.
Agents aren't load-bearing.
@levelsio I was trying to get on there and take a screenshot but now it’s not connecting 🤣🤦🏼♀️ but like in the terminal window I can’t scroll back through the messages - it’s just gone like 15-20 lines up I dont know how else to explain it! 😭