Korva is a scroll of the internet’s best ideas.

Agents push Git too hard

Coding agents are probably the most popular kind of agent, and for the same reasons they understand the file system, they understand Git. They almost understand it too well. With loops, any wasted time is expensive, so rate limits and downtime hurt that much more.

An agent-native alternative

That has people asking what a more agent-native, Git-like system would look like. It could serve all kinds of agents, because most agents people build are effectively coding agents. The model is usually the same and the set of tools is very similar. The task just might not be producing a software artifact.

Beyond coding agents

Picture a lawyer working with an agent, a customer service person working with an agent, or an accountant working with an agent. In each case the agent is working inside a file system and a Git-like construct. Nobody has yet found the exact product that fits.

How does Git store history so compactly?

Git stores each file version as an object named by a hash of its contents, so identical content is saved once. Periodically it packs objects together and stores many of them as differences from similar ones. Linus Torvalds wrote Git in 2005 to manage Linux kernel development. The Linux kernel repository, with decades of history, remains one of its best-known uses.

Why would agents strain a Git host?

GitHub documents rate limits on its API, which cap how many requests an account can make in an hour. A person might make a handful of pushes and fetches in a day. An agent can run many tasks in parallel, each one cloning, committing, and pushing. Those requests add up against the same limits.

Why do parallel agents create merge conflicts?

Git merges two branches by comparing each against their common ancestor. When both changed the same lines, it cannot choose between them and marks a conflict for someone to resolve. Several agents editing one repository raise the odds that two of them touch the same lines. Each conflict needs a resolution before the work can be combined.

More from Lightspeed Venture Partners

Tech & AIAgents work best in foldersThe ideal environment for an AI agent is a file system, with files and folders and commands to work with them, because models learned from humans using computers that way.
Enjoyed it?