Brian Balfour

<start of post>

I’ve spent the last decade obsessed with how teams build. From the early days of Reforge to the scaling years at Underdog, the question was always the same: how do we move faster without breaking the things that matter?

But the last six months changed my perspective entirely.

Watching how AI agents are actually being integrated into the workflow—not just as chatbots, but as core contributors—has been a whirlwind. It’s the most energized I’ve felt since the early days of mobile.

So I’m leaning in.

AgentFlow is a lightweight framework for mapping AI-human handoffs in product orgs.

Before we started building it, I asked a few CPOs how they were managing AI output quality.

Most told me they were just "vibing it" and hoping for the best.

That answer bothered me. It felt like we were repeating the mistakes of the early agile days.

So we started documenting the inflection points where human oversight actually adds value vs. where it just creates a bottleneck. Three things surfaced:

The first was the "Context Gap." Agents are great at execution but terrible at understanding the 'why' behind a PRD. We found that a 5-minute human context sync saved 2 hours of debugging.

The second was the "Feedback Loop." Most teams were treating AI as a black box. By opening up the intermediate steps, we saw shipping velocity increase by 40%.

None of these insights came from a textbook. They emerged from building in public and breaking things in real-time.

That’s the goal of AgentFlow: to move from "vibe-coding" to a structured system for the next decade of software.

It’s early, it’s rough, and it’s definitely not a finished product.

Please fork the repo, tell me what I got wrong, and share your own workflows. Link in the comments.

Come join us. We’re just getting started ⚡⚡⚡

Sarah, Mike, Elena, David, Chris, Julian, Amit, Raymond, Doug, Rob, Michael, Marco

<end of post>