What version control looks like when AI agents write the code

When a thousand engineers spin up a hundred agents each, Git's merge request model breaks at machine speed.

Git was built for human-to-human collaboration, but what happens when a thousand engineers each spin up a hundred AI agents? In this episode, Lee Faus, CEO and founder of Atomic Software, joins the We Love Open Source podcast to share why version control needs to evolve for the age of AI agents, and how petri nets can transform the way developers handle pipeline failures.

Lee brings decades of experience with version control systems, from ClearCase and CVS to Visual SourceSafe, Subversion, and Git. That journey led him to discover Pijul, an open source project built by a French researcher proving patch theory through his doctorate thesis. Four years of conversations and watching GitLab customers struggle with agentic workflows convinced Lee it was time to build something new. The problem became clear: When enterprises run thousands of merge requests from AI agents in short periods, Git’s human-to-human design hits its limits. Atomic is built from the ground up with AI and agents in mind, rethinking how version control should work when machines write code at machine speed.

Read more: Open source for the age of agent-native development

Lee explains software development through an hourglass effect. Teams now spend more time upfront building context, prompting AI, and configuring skills and tools. Then syntax writing, which used to require formal training and deep language knowledge, happens instantly. But the other side of the hourglass presents a new challenge: Validating that what you requested is what you got. When a thousand engineers generate a hundred changes per day, traditional review processes break down. Atomic addresses this by storing AI provenance and creating change records that work fundamentally differently than Git. Lee warns that developers coming from Git need to approach Atomic with fresh eyes, comparing it to telling an AI to ignore everything discussed before and start over.

The companion product, Circuit Breaker, tackles another pain point in modern development. Built on petri nets, a state modeling system that goes back decades, Circuit Breaker lets developers run mini pipelines against every change. Lee uses a relatable example: Imagine a 20-step pipeline where step 19 fails because your Kubernetes manifest was wrong, but your Docker image and everything before it was perfect. Traditional pipelines force you to re-run all 20 steps. Circuit Breaker breaks jobs into states, so you can fix the manifest, apply it to your existing Docker image, and continue from that point. You can stop, restart with the right precursor, or drop everything and start fresh. It’s the kind of optionality current pipelines don’t offer.

Key takeaways

• Version control needs to evolve for AI agents: Git was designed for human collaboration, but when enterprises run thousands of agents generating massive merge request volumes, the system shows its age.
• Petri nets enable smarter pipeline management: Circuit Breaker lets developers restart pipelines from failure points instead of re-running entire workflows, saving time when issues occur late in the process.
• The hourglass effect shifts where developers spend time: AI collapses syntax writing to near-instant, but validation becomes the new bottleneck. Teams need tools that handle provenance tracking and change validation at scale.

Lee’s projects live on GitHub under the atomic.dev organization, where developers can explore Atomic, Circuit Breaker, and documentation. His advice? Approach these tools with fresh eyes rather than trying to map Git concepts directly. The shift from human-speed to machine-speed development demands new thinking about how we track, validate, and manage change.

More from We Love Open Source

• 15 open source backup solutions to protect your data
• Open source for the age of agent-native development
• Stop guessing, start measuring developer engagement
• 6 must-read Linux and open source tutorials
• 5 forces driving DevOps and AI in 2026