Reliable finance agents: keep control flow deterministic, let models only judge

AI Engineer · youtube · 2026-10-13

Karthic Rao of Perpetual uses a wealth-management example to show why "fill out a form for Mr. Harper" becomes a multiday process: identify the right household, find missing information, contact the client, wait, follow up, obtain a signature and retain evidence. A fixed flowchart breaks on ambiguity, while an improvising agent creates a different reliability problem.

His design puts control flow, evidence and external actions in a deterministic system; models handle interpretation and judgment at explicit, typed points. A judgment can create a smaller plan, but that plan must pass the same compilation and approval boundaries.

The planner turns a request into a semantic graph whose nodes point back to the request; a compiler validates it against the firm's action contracts, returning a valid plan, a clarification request or a refusal.
Approved plans are immutable; external actions have separate gates tied to the exact payload. Waits and follow-ups are state-machine steps, and uncertain action outcomes go to manual review rather than automatic retries.
In an illustrative 20-decision run, 19 decisions go to a smaller model and one to a frontier model; that ratio is a design target, not a measured fleet result.
Typed, logged judgments open a path to privately hosted inference and future training, though the training loop is still forthcoming.

Original post →

Reliable finance agents: keep control flow deterministic, let models only judge

More from AGI Musings