The Loop Dispatch
Issue 006 · March 14, 2026
Lead essay · Quality Engineering
Most QA functions were built for a world where humans wrote all the code. That world is gone. Here's the operating model we've deployed across 30+ engagements to replace it.
Most QA functions were built for a world where humans wrote all the code. That world is gone. Here's the operating model we've deployed across 30+ engagements to replace it.
About the Author
Ben Fellows
The fundamental problem with traditional QA isn't headcount, tooling, or process. It's the operating model itself. QA was designed as a gate…
All Articles
Most QA functions were built for a world where humans wrote all the code. That world is gone. Here's the operating model we've deployed across 30+ engagements to replace it.
Ben Fellows
8 min read
Everyone is shipping AI agents with giant prompts. The teams getting real leverage are building purpose-built pipelines with roles, permissions, and project-specific rules instead.
Marcus Rivera
6 min read
We redesigned the TDD workflow for AI-assisted development. Tests define intent, AI accelerates implementation, and engineers own verification. The results changed how we think about quality ownership.
Sarah Chen
7 min read
Velocity metrics tell you how fast you're moving. Leverage metrics tell you whether the speed is sustainable. Here's the measurement framework we use with every client.
Ben Fellows
5 min read
Most maturity models are academic exercises. This one was built from 30+ real assessments and tells you exactly what to fix first based on where the leverage is highest.
Priya Sharma
6 min read
The SDET role is being redefined. Manual test execution is gone. What remains is automation engineering, infrastructure design, and the ability to operate AI-assisted workflows. Here's how we hire for it.
James Okafor
5 min read
How readers use the writing
The essays are most useful when they give you language and evidence for an argument you already wanted to make.
“I sent “Coverage Is a Vanity Metric” to my CTO. Two weeks later we redefined our QA OKRs.”
Series-B fintech · ~50 engineers
Used: Coverage Is a Vanity Metric
“The AI-generated tests piece reframed our vendor review. We rewrote the evaluation rubric the same week.”
Healthcare SaaS · 120 engineers
Used: Why Your AI-Generated Tests Are Worthless
“I forwarded the staffing-model essay to our head of engineering. It started a conversation we'd been avoiding.”
E-commerce platform · 200+ engineers
Used: The Staffing Model Is Broken
Names + companies anonymized at the speakers' request.
Watch · From the desk
Apr 28, 2026
If you want to start controlling AI-generated code today, this is the simplest way I’ve found to do it. In the previous videos, I talked about why agentic development breaks at scale and introduced the concept of policy as code as a way to fix it. In this video, I’m showing how to actually get started. The idea is straightforward. Instead of relying only on prompts, rules, or memory to guide AI, you introduce a deterministic layer that scans your codebase and flags violations. Think of it as a much more comprehensive, fully customizable linting system that works alongside tools like Claude. What surprised me is how easy it is to get a first version working. In this walkthrough, I show how you can go from zero to a basic policy as code setup in a very short amount of time. We start by generating a small set of rules, wire up a simple scanner, and immediately run it against a real codebase. Even with a basic setup, you’ll start catching issues and inconsistencies right away. This is not the full system I use in production. At scale, this turns into hundreds or even thousands of rules, with more advanced concepts like evidence layers, caching, and reporting. But the goal of this video is to show that you don’t need any of that to begin. If you’re using AI to write code and you’re starting to see drift, inconsistency, or quality issues over time, this is a practical way to start putting guardrails in place. Over time, what I’ve found is that as you add more rules, the amount of drift drops significantly, and the system becomes more reliable without slowing development down. If you haven’t watched the earlier videos in this series, I’d recommend starting with those for more context on why this approach exists and how it fits into a larger agentic workflow. If you try this yourself, I’d be interested to hear what kinds of rules you end up writing and what it catches in your codebase.
Apr 27, 2026
I spent time building with “agentic factories” - multi-agent pipelines that promise fully autonomous workflows. On paper, they look like the future. In practice, they broke down in ways that matter: reliability, coordination, and real-world constraints. In this video, I break down where these systems failed, why they fail structurally, and what actually worked instead in production. If you're building with AI agents, this will save you time (and probably some pain).
Apr 24, 2026
Claude and other AI agents are incredibly good at writing code. The problem is they don’t stay consistent over time. In the first few iterations, everything looks great. Output is fast, patterns are mostly correct, and it feels like you’ve unlocked a new level of development speed. But as the codebase grows, small inconsistencies start to compound. Patterns drift, structure degrades, and eventually the system becomes harder to maintain than it was before. That’s the problem this video is about. In this walkthrough, I break down how we use a concept called policy as code to control AI-generated code in real systems. Instead of relying only on prompts, rules files, or memory, we introduce a deterministic layer that enforces how code is allowed to be written. Every time an agent makes changes, those changes are checked against a large set of rules. If something doesn’t match the expected patterns, it fails. The agent has to fix it before moving forward. This ends up acting like a much more comprehensive version of linting, but tailored specifically to your architecture, your patterns, and your codebase. The result is that we’re able to keep the speed benefits of AI while dramatically reducing drift and long-term degradation. This video focuses on how the system works in practice. What kinds of rules we write, how they’re structured, and how they integrate into an agentic workflow using tools like Claude. If you’re experimenting with AI coding and running into issues with inconsistency or quality over time, this is one approach that has worked well for us. I’ll also be doing follow-up videos on how to implement this from scratch and how it fits into larger agentic pipeline systems. If you’ve tried something similar or have different approaches to controlling AI-generated code, I’d be interested to hear about it.
Every article is backed by a published methodology paper. Download the frameworks, run the assessments, and see the full system.
90-Day QA Leverage Plan
Use it on Monday · Editable doc + Notion template