---
title: "Front-line expertise, then and now: the review agent found what the builders missed"
url: "https://toddpaulbrownjr.com/writing/front-line-expertise/"
author: "Todd Paul Brown Jr."
description: "Weekly trainings at NP Accel and an adversarial review agent are the same mechanism: a channel for front-line expertise with no stigma for being wrong."
kind: "article"
updated: "2026-09-26T02:50:30+00:00"
---

# Front-line expertise, then and now: the review agent found what the builders missed

The people or systems closest to the work catch what the builders structurally can't see in their own output, but only if there's a real channel for what they find and no stigma in being wrong. The leader's job is building that channel.

I've seen this principle work on both sides of the agent divide. When I ran marketing operations at Neil Patel Accel in 2019 and 2020, I started weekly training sessions where we dissected and rebuilt marketing strategies together. The sessions weren't meetings; they were working sessions with written follow-up. I oversaw scheduling, led the meetings, and made sure what we found turned into training material and processes the fulfillment team could actually use. A reporting system I built there was later adopted company-wide — it worked because it was built to spread.

In August 2026, I pointed a five-agent review pass at a generative audio app I'd just finished building. The brief was simple: find what's wrong. The agents had independent web research enabled and no instruction to be polite about it. They found two showstopper signal-processing bugs that the build agents, the test suite, and my own listening had all missed.

The first bug: stereo panning did not exist. The equal-power pan law used `L = cos((1−p)·π/2)` and `R = sin(p·π/2)`. By a trig identity, those produce the same number for any pan value `p`. Every pan position in every Potion produced identical gain on both channels — mathematically perfect center imaging, every time. The second bug: the "master limiter" was not a limiter. It was a waveshaper with no threshold, no envelope, and no unity region, applying constant audible distortion to every sample at every level. Both were fixed the next day with a real equal-power pan law and a real peak limiter with threshold and attack/release envelope.

Every automated test had passed throughout the build. The tests verified the code agreed with its own contract; they did not verify the contract matched what the product needed to be true. A human listening to binaurally-rendered procedural audio on cheap earbuds in a loud room will not catch a pan law that self-cancels, especially when the rest of the signal chain is working. The review agents, briefed only to find problems and armed with reference implementations from the web, caught both within a few hours.
The counter-case on the agent side: a lead agent I briefed to coordinate research work launched ten subagents at once with no budget policy. The session burned through a five-hour budget in 33 minutes because nothing coordinated them. Eight of ten subagents were killed mid-write. Unstructured input burns the shared budget just as fast as too many voices with no process. The channel needs structure, not just permission.

I've been thinking about an "I told you so" board for settling disagreements — design a test, run it, post the winner and loser publicly. The idea is that it starts with me being proved wrong and posting that outcome myself, then inviting others to join. Every result logged, wins and losses, so being wrong carries no stigma. I also want to start crediting technicians for their creativity when they generate a lot of hypotheses, because people don't often associate creativity with data analysis and I want to change that. Those are ideas I described in 2025; I haven't run them yet.

What carried over from the NP Accel sessions to the adversarial review brief is the same mechanism: a defined place for the people (or agents) closest to the work to say what's wrong, with written follow-up and no penalty for surfacing a problem. The weekly session turned findings into training material and fulfillment processes. The review brief turned findings into fixes the next day. Both worked because there was a real channel and a commitment to act on what came through it.