ampm: A software team that ships reviewed pull requests

Claude Code changed everything for me.

The leaps in what we can build today compared to just a few months ago are unbelievable. I've built side projects that had been stuck in my head for years. Things that once felt like enormous undertakings suddenly became weekend projects.

But every technological breakthrough comes with a moment of freedom, and then you discover new boundaries at the next level of scale.

For me, that boundary was managing agents. I was juggling multiple Claude Code sessions, constantly switching contexts, checking progress, answering questions, and figuring out what to do next. The coding had become almost effortless, but managing the coding was becoming a full-time job.

I wanted something different. I wanted Kanban. I wanted cloud-hosted agents working autonomously, while I checked in, answered questions, and gave direction on my own schedule. The popularity of Vibe Kanban made me realize I probably wasn't alone.

But I wanted to go further.

I wanted agents that were properly sandboxed. No giving an AI the keys to the castle. Agents should be able to read code, research, write code, and commit changes. Everything else—approval gates, GitHub interactions, workflow transitions—should be enforced by deterministic code, not left to an LLM's judgment.

I wanted to use my existing Claude and Codex subscriptions. I wanted Codex reviewing Claude's work, or vice versa. I wanted automatic fallbacks when usage limits were hit, so the work could keep moving. The spice must flow.

And above all, I wanted to stop babysitting agents.

I tried n8n, Kandev, Paperclip, Multica, and a bunch of other tools. Each had interesting ideas, but none quite delivered what I was looking for.

So I did what any self-respecting developer with newfound AI superpowers would do: I built it myself.

That's how ampm was born.

The philosophy is deliberately opinionated. Software development already has a well-established workflow: plan work, implement changes, review code, open PRs, merge. I didn't want a visual workflow builder or an infinitely configurable automation engine. I wanted a system that understands software engineering and just gets on with it.

And most importantly, it's built around asynchronous collaboration, not synchronous chat sessions. Your agents work while you're doing something else. They come to you when they need input. You decide when to engage.

Today, I routinely run 10–25 concurrent agents, checking in on them when it suits me. It feels like another enormous leap in what's possible.

But building ampm also taught me why this problem is so much harder than it looks.

Agents sometimes modify files outside their assigned worktrees. PRs get merged, closed, or rebased outside the system. GitHub state changes unexpectedly. Agents get stuck, fail, or run out of quota. And when you've got dozens of tasks in flight, even grooming, deduplicating, and prioritizing your backlog becomes its own job.

A demo that launches a bunch of agents is easy. A reliable system that coordinates them, survives failures, and keeps moving without constant supervision is much harder.

That's ultimately why I'm sharing ampm.

I've spent 15 years in software engineering, much of it as a tech lead. I've learned a lot about organizing engineering work, managing dependencies, designing processes, and keeping projects moving. I wanted to encode that experience into a tool that other developers could benefit from, without having to spend months building and debugging their own orchestration systems.

Claude Code gave us an incredible productivity leap. Now, ampm to unlocks the next 10x.

Ideation, design, whiteboarding, voice dictation, gathering signals from users, analyzing funnels, synthesizing interview notes, and automatically turning insights into actionable work are all on the horizon.

My vision is an entire software engineering team at your fingertips. Not just agents that write code, but a system that helps you figure out what to build, why to build it, and how to ship it.

I built ampm because I desperately wanted it to exist.

Now I hope it gives you the same feeling of freedom that Claude Code gave me when I first tried it.

Check it out at ampm.dev. I'd love to hear what you think.