A Day Managing AI Agents (Part 1)

I still remember the first day I used ChatGPT. It led to a weekend of discovery, awe, and a newfound ability to round out my coding knowledge, creating apps in the span of a weekend. Since then, I’ve been waiting for another “aha” moment like that, one of those experiences that makes you feel like you’re getting a glimpse of the future.

Well, last week, it happened again.

Ermioni Clemence and I met in the office with a mission: create a prototype for an app for Dan Diasio. We changed our Teams status to “vibe coding, forgive delays in responses,” and we got to work.

This time, not with an AI chatbot, but with Replit.

Replit still has a chat interface, but instead of one LLM, it gives you access to a team of specialized agents, from architecture to backend development. It felt similar to Anthropic’s Claude Code, but more structured. And it sparked a very different workday, maybe a closer resemblance of what a “normal” working day could look like in the future.

Here’s how it played out:

🧠 9:00–11:00 — Brainstorming the “why”

Emmi and I started with ideation: What problem were we trying to solve? How did we want the user to think, feel, and respond? What should the UI/UX look like?

🧾 11:00–12:00 — From whiteboard to prompting

With our whiteboard filled out, we snapped a photo and uploaded it into one of our Custom GPTs (already trained on a lot of our previous work). It summarized our concept and created a technical brief:

“Include all the information a development team would need to create the MVP.” We refined it, added details and structure, then jumped into Replit.

🤖 12:00–12:30 — Agents take the wheel (and we… walk)

Uploading the brief was basically the whole setup. A team of agents started breaking down our big ask into features and user stories, while the orchestrator started assigning work, without us.

Emmi and I looked at each other with the same question: “Now what?” From earlier test runs, we knew the process would take ~15 minutes. Perfect timing for a walk.

During the walk, we brainstormed what we’d want to add in the second run—what we forgot to specify in the brief. After our first lap around the office, we peeked at the monitor: the architect agent had hit an issue. We decided to give it time. “Let’s do another lap.” Another lap later, the issue was resolved, on to the next feature. Another lap, more ideas. Our mental backlog kept growing. And then… the first run was done.

✅ 12:30–1:00 — First build review + next prompt

The app was there—pretty much exactly how we described it. It executed the details of our vision, and it got creative where we lacked nuance or structure. Our walks gave us space to reflect, and within moments we captured our changes: more features, UI improvements, refinements… and started another run.

“I guess we have time for lunch?”

🍽️ 1:00–2:00 — Guilt-free lunch

Ermioni and I both have (slightly) stakhanovite tendencies. We’re guilty of eating lunch at our desks—either getting work done or going camera-off on calls while we sneak in food.

This time was different. We knew work was getting done anyway.

It was guilt-free freedom: time to catch up, and time to keep thinking about what we were building. We came back just as Replit wrapped its second run. It was better—it implemented most of our feedback, but not all of it.

That’s when we learned something important: the first development run is different from the ones that follow. You can start big (long prompt, lots of features), but as your vision sharpens and your requests get more demanding, you often have to go one feature at a time.

🏃 2:00–4:00 — The loop: ideate → direct agents → walk → test

Ideation, give the agents direction, take a few laps around the office.

We were on a roll. It was a new routine—different from any day we’d had on the job. On our walks, we ran into coworkers and friends. They followed us back to our room and tested what we were building incrementally.

Over a couple hours, we collected more feedback and user testing than we’ve ever gotten on Day 1 of a development project. And finally—ten runs and ten thousand steps later—we had our app.

🧪 4:00–5:00 — Forks and parallel builds

“Is this really what we’re trying to build?”

We were presenting to Dan in an hour. Usually, questioning your work in the last hour before presenting to your boss is… not ideal. But this time, we weren’t alone.

We knew our agents could handle a lot in one hour. We started thinking through variations: different user experiences, different “so what’s.” We forked the project a couple times, and suddenly we had several teams of agents simultaneously working on separate builds—ready just in time.

🎤 5:00 — The reveal

“You did THIS today?!”

Dan smiled as we walked him through a few variations. He knew what he was doing when he got access to the coveted Replit licenses—but the result was still surprising.

We mutually agreed:

App 1 was sh*t. App 2 was a fun idea, but not worth scaling. But App 3… that was headed in the right direction.

We aligned on next steps: keep building for the next week, with the goal of deploying a ready app in the next few business days.

A few things happened on a cold, snowy day in Chicago. Ermioni and I accomplished something in a day that would normally take a full-stack development team a handful of sprints. We also had one of the most collaborative days we’ve ever had—reinvesting the time our agents saved us into more brainstorming, planning, and collecting feedback.

And we ended the day with clarity: where we were headed, and the confidence we could ship. Also… my Oura ring was pinging me with joy because I hit all my movement goals.

This was just one day. But sometimes one day is all you need to get a glimpse of the future.

I saw a future where ideas and creativity become more valuable than execution skills. A workday with faster collaboration and tighter feedback loops than ever. We didn’t feel replaced—we felt like we did our best work, freed from constraints like time, development skill depth, or the funding required to staff a team.

We’re entering an age where most people will use AI to do things differently—but the real winners of tomorrow are the ones who leverage AI to do different things.