본문 바로가기
Case · 2026.10.12

Leave control flow, evidence and outbound actions to deterministic code and confine the model to typed decision points — ambiguous requests no longer break the workflow, and compliance can sign off on it

Reliable, Smart Workflows for Finance, Wealth and Insurance — Karthic Rao, Perpetual — AI Engineer · Read the original ↗

This page is a translation of the Korean original. Read the Korean original →

What you get

  • The flow holds even when a vague one-sentence request grows into a process that runs for days
  • Every outbound action passes an approval gate bound to its exact content
  • Decisions are recorded as types and logs, opening a path to privately hosted inference

When to use it

Apply it when automating long-running processes in regulated work such as finance and wealth management, where policy breaches are costly and the process mixes resolving ambiguity, multi-day waits, follow-up reminders and preserving signature evidence. It also assumes a situation where sensitive data forces inference onto privately hosted models.

How to do it

  1. Draw the boundary so control flow, evidence and external actions sit in deterministic systems, and the model handles only interpretation and judgment at explicit, typed points.
  2. Have a planner translate the request into a semantic graph whose nodes trace back to the wording of the request.
  3. Have a compiler validate that graph against the company's action contracts and return only one of three results: a valid plan, a request for clarification, or a refusal.
  4. Freeze approved plans as immutable, and give each outbound action its own gate bound to the exact payload.
  5. Handle waiting and follow-up reminders as state-machine steps, and send side effects with uncertain outcomes to manual human review instead of retrying automatically.

What to borrow

You can apply the same split to your own agents: let code, not the model, decide what runs next and what goes out. The parts to copy directly are receiving planner output only as valid, question or refusal after checking it against a contract list, and routing calls with uncertain outcomes to a human review queue instead of retrying automatically. It is safer to treat the figure for the share of small models as a design target only.

Background — what the original covers

Karthic Rao, CTO of Perpetual, presented a design for automating work in regulated industries. His starting point is that in wealth management a single sentence such as "Fill in a form for Mr. Harrison" actually becomes a process lasting days: identifying the household, collecting missing information, contacting and waiting on the client, sending reminders, getting signatures and archiving evidence. The problem he frames is that fixed flowcharts stall on ambiguity, while agents that improvise create a different kind of reliability problem.

#Agents#Safety

← All practitioner tips