Why we're writing this
A blog about the unglamorous half of AI engineering: what happens when an agent proposes an action that a person has to stand behind.
Most of what gets written about AI agents is about making them capable. Very little is written about what happens at the moment an agent proposes to do something a person has to stand behind — issue the refund, send the email, file the claim, deploy the model.
That moment is where we spend our days, and it turns out to be much stranger and more interesting than "add an approval step." So we're going to write about it here.
What this blog is about#
Concretely, the problems we keep running into and the ones our design partners bring us:
- Approval design. What a reviewer needs to see before they can honestly say yes. Why a bare approve/reject button is a liability rather than a control.
- Policy. How to express "a human decides this" as a rule rather than a hard-coded
if, and how to keep an AI from quietly deciding it doesn't need one. - Routing, quorum, and escalation. Who gets asked, how many of them, and what happens when nobody answers. Silence is a design problem, not an error case.
- Provenance. What has to be recorded so that, a year later, someone can reconstruct not just who approved something but what they were looking at when they did.
- Workflow engines. What Inngest, Temporal, Trigger.dev, and Hatchet give you, where their primitives stop, and why that boundary sits exactly where it does.
Some of it will be opinionated. Some will be things we got wrong and had to change. Where we haven't shipped something yet, we'll say so — the same way we do on the rest of this site.
Why we think it's worth your time#
Human-in-the-loop is treated as a checkbox on a compliance form, and it deserves better. The difference between an approval system that works and one that decays into a rubber stamp within a quarter is a hundred small design decisions, almost none of which are written down anywhere.
We are making those decisions in public, on a product that lets agents move real money. We'd rather share what we learn than pretend we arrived at it obviously.
If you're building something similar, disagree with us loudly. That's the point.
Ratifia pauses an AI workflow for a human decision, gives the reviewer the full context, and resumes with the verdict. It runs on top of the workflow engine you already use and never executes your code. Request early access.