managent

It started at a hackathon. I was building with other people, and we were all using AI to write code. When we tried to merge our AI-written branches into the same repo, we hit hella merge conflicts. Files overwriting each other. Work that looked fine on one laptop breaking the moment it landed next to someone else’s. The agents were smart. Working together as a team of humans + agents on one codebase was the hard part.

That stuck with me. So I got into researching the problem and its scope. With co-authors, I studied 33,596 agent-authored pull requests across 2,807 repositories. The paper is on arXiv: AI Agent Pull Requests on GitHub: Frequency, Structure, and Merge Conflict Rates.

What we found was hard to ignore. When you look at exact temporal overlap, 40.2% of repositories already have co-active agent-authored PR pairs — and those pairs account for 79.4% of all agent PRs. Widen the window to a week and it jumps to 53.4% of repos and 95.0% of PRs. Cross-agent pairs conflict at 41.7% vs 19.8% for the same agent working alone. Coordination, not codegen, is the missing layer.

And so I started building managent — which is what I’m working on today.

Managent is one control plane for AI coding agent fleets — Slack for AI agents. Think of it like a shared calendar for code: before an agent starts on a file, it books that territory — everyone else waits their turn automatically. You can’t collide on territory you’ve leased.

The unique value prop is doing this at enterprise scale: different devices, different agents, and large monorepos on one map. An engineer’s laptop, CI, and the cloud. Claude next to Cursor next to Devin. Thousands of files where hotspots get slammed. Agents lease territory before they write, queues stay visible, and every merge is replayed and verified before it ships.

Within respected leases, we’ve seen 0 conflicts in measured merge-tree replays. The inherited baseline for uncoordinated concurrent pairs is still about 1 in 5. That’s the gap we’re closing.

We’ve spent years asking how to make a single agent smarter. That’s useful — and it’s also incomplete. Intelligence without coordination doesn’t compound. It collides.

Ninety percent of developers now use AI coding tools. Seventy-four percent have adopted a specialized coding agent. The bottleneck is no longer whether an agent can write code. It’s whether a whole team — humans and agents, across devices and vendors — can share a large monorepo without quietly breaking each other’s work.

That’s a systems problem, and it’s the same shape as my broader thesis: the interesting work isn’t always making people or machines more intelligent. It’s designing the layer that makes the intelligence we already have useful at scale. Managent is that layer for agent fleets at the enterprise level.

reach me at: nithilan@managent.tech

learn more at: www.managent.tech