Who Manages the Agents?

Welcome to this installment of Veo, where every week, I share a bit about what myself and others are seeing on the cutting edge of software product development.

In this issue…

Did someone forward this to you? If you like it, you can sign up here.

Want to grab time on my calendar? You can do that here.

-John Cox

I learned FORTRAN the summer between my sophomore and junior year of college.

That's the only formal programming class I ever took.

But my teams and I have been building software for 20 years across financial services, healthcare, retail, IoT infrastructure, cloud systems, and everything in between. So many different languages, different frameworks, different platforms, different integrations. A career spent learning on the job.

We are now taking that knowledge and experience (minus the FORTRAN) and building the Virtual Engineering Organization where we manage a team of AI agents that produce software faster and better than all the engineering teams I've worked with (sorry y’all). We’ve been developing the tools and processes to help us deliver projects for our customers, and I’ve sat in the operator’s (that’s the human that leads the agents) seat many times now.

So here's what I've learned about what it actually takes to be a world-class agentic engineering operator.

You don't need to be a professional software engineer. You don't need a CS degree. You don't need to be able to write a perfect SQL query from memory. Sure these things can’t hurt you, but they’re not required.

But I believe you do need three specific capabilities.

First, you need to understand the breadth of what you're building. That means knowing the business objectives, the user needs, and the technical architecture well enough to recognize when something's on track and when it's not. You're not writing the code, but you need to know what the system is supposed to do and how all the pieces fit together.

Second, you need pattern recognition. When you're reviewing code or architecture, you need to be able to spot when something feels off—even if you couldn't fix it yourself. That's not about deep technical skill. It's about experience across enough systems to develop the intuition to know when to step in.

Third, you need to be able to coach and guide. This is a management skill, even though you're interacting with agents instead of humans. You need to know when to redirect, when to push back, and when to let the work proceed.

This last one matters more than people realize.

While helping out on a project I got excited about some of the features the agents suggested—like a military branch theming system that sounded cool in the moment. Of course I said “of course!” but it ended up derailing what I was working on and created a lot of drag later in the project when other agents didn't use it correctly.

That's scope creep, and you need to recognize it just like you would with any team.

On the flip side, you also can't just blindly trust the agents. That leads to slop—code that technically works but doesn't follow the guidelines or intent of the system. And you can't distrust them completely either, or you're not getting any efficiency gains.

Trust but verify.

The irony is that working with agents isn't that different from working with today’s human teams—especially when most of your interactions are over Slack and email anyway.

There are actually some advantages. No social pressure. No feeling like you're asking a dumb question. The agents have broader access to knowledge about frameworks and languages than any single human could. And they work when you want them to—even when you're not working.

You don't need to be an AI expert. You need to be an effective operator.

Three questions to assess your (or your team’s) readiness:

If you answered yes to those three, you're already 80% of the way there. The rest is learning the agentic SDLC—and we're actively refining that process with every project we run and every day we build the system.

We're not trying to replace humans. We're trying to empower people to build better software that's more aligned to specifications, higher quality, more consistent, and enterprise-grade.

The operator role is the linchpin.

If you're a senior engineer, an architect, a technically-adept product manager, or a product designer who understands systems—you probably fit the bill.

More links! Here are a few articles that I came across that caught my attention. See what you think.

1. OpenAI announced that they will start introducing ads in the free and Go tier of ChatGPT. So we’re basically back to Google? I’m already leery of OpenAI because of what I see as shiftiness. I happily pay for my AI services - I have two paid Claude accounts - because the return is tremendous. Maybe in some ways they are a victim of their own success with acquiring users onto the free platform. It just seems like the objectivity (or at least the perception of it) is at stake once advertising is involved.

2. Viral adoption of Claude Code is putting heat on software companies’ stock prices. Sorry, paid site (WSJ). I guess Claude Code is so good that people are second-guessing all those expensive SaaS licenses (I know I do every day).

3. Hard reps make you grow. From Arnold’s Pump Club newsletter, the idea is that easy reps do way less for you than hard reps. Failure is the best teacher in business and turns out the same is true for muscle growth. I read this both literally (hey this can help at the gym) and in a way that applies to our business (keep trying to solve hard problems and feel the burn). And that’s the second time in a few weeks that I’ve referenced this newsletter.