# Marcus Trevino, Head of Support Engineering at Fieldwire (SaaS, ~170 employees) — read of Support Agent Guard, June 16 2026

> 9 years in support infrastructure, currently running a hybrid team where our Intercom bot handles tier-1 and three humans handle everything else. We started routing refund requests through a Claude agent in March. It has already scared me twice.

## How I got here

Searched "ai agent guardrails production" after a Monday incident where our Claude integration issued a $400 refund for a plan we'd already cancelled. The customer got the money, didn't exist in our active accounts, and no one on my team knew it happened for six hours. I've been looking for something that sits between the agent and the action ever since. This page was the fourth result. I clicked it.

## What I clicked first

"It approves refunds without checking inventory. It accesses payment history without permission. It makes commitments your team can't keep."

That's the exact list. That's not a marketing team guessing at pain points, that's someone who has sat in a post-mortem about one of these. I read that paragraph twice.

## Where I paused

"We don't have live customers on this idea yet."

I had to re-read the whole page after hitting that line. Because up to that point I was reading this as a product. Feature list, integration claims, "Deploy in 15 minutes" -- all of it reads like a product. Then the bottom of the page tells me it's... an idea for sale? A strategy package? I'm buying a dossier and a code starter, not deploying software?

I'm still not completely sure what I'd actually get if I clicked "Adopt the build" for $99. A starter repo? A Notion doc? Something I'd have to finish building myself? The page doesn't answer that, and it's a big question.

## What I distrusted

"Deploy in 15 minutes." I've been deploying things for nine years. That line gets written by someone who has never had to route an action interceptor through a company's change management process, get it approved by security, and coordinate the release window with three other teams. The technical integration being fast doesn't mean deployment is fast. That claim makes me trust the other claims less.

Also "Your agent learns" -- this appears once with no explanation. Learns what? Learns which actions are risky? From approvals and rejections? Via fine-tuning? Is this a hallucination of a feature or is it real? When I see a phrase like that with no follow-up, I assume it's not real yet.

The $-24,672 year-one Fermi number is also confusing. Negative take-home means this loses money in year one? I assume that means build costs vs revenue, but the page doesn't explain the math. I almost scrolled past it as a bug.

## What would convince me

A 90-second screen recording of someone deploying this into a LangChain agent and then triggering a risky action. I want to see the intercept happen in real time and watch a human approve it in the UI. I don't need a polished demo environment -- a messy terminal and Slack notification would actually be more convincing than a slick deck.

Beyond that: one sentence explaining what I'm actually getting. "You receive a working Node.js middleware layer, an admin dashboard you self-host, and a policy config file." Something that concrete. Right now I can't picture it.

## What I'd ask in an email reply

1. When you say "drop-in integration," do I get a repo I modify, or a hosted service I call via API? Because those are two completely different security conversations I'd need to have with my team.

2. What happens to the action request while it's waiting for human approval? Does the agent block, timeout, or return a holding response to the customer? That's the UX cliff I'd fall off in production.

3. You mentioned "your agent learns" -- is that live functionality or a roadmap item? I'd rather you just not say it than say it and mean "coming soon."

## Verdict: on-the-fence

The problem is real and the page names it accurately, which is rare. But I can't tell if I'd be buying a product or commissioning someone to help me build one, and that confusion is the page's fault, not mine. If there's a working interceptor I can actually run, I'd reply today.

---
*Memo by skeptic persona, generated 2026-06-16. Studio breaks own self-grading loop.*