2026-06-10 · Grit: rewriting Git in Rust with agents

Transcript

Transcript. Paraphrased; sources in notes.md.

HostIt is June 10th, 2026. Welcome back to the show. Today we are looking at a trio of threads from Lobsters that dive into how we build things. We have a massive attempt to rewrite Git, a philosophical look at why we build side projects in the first place, and a Wikipedia guide on how to spot the trail left by LLMs. Samantha, good to see you. You look like you have already had your third espresso.

GuestThree is accurate, Daniel, and it is entirely necessary. I was actually just clicking through that Git rewrite story. The idea of using an agent swarm to tackle something that complex, instead of manually chasing down every pointer, is basically my dream workflow. It makes the prospect of building internal tools for my team feel so much faster. I am ready to dig into this.

HostLet us start there. Scott Chacon has launched a project called Grit. It is a Git implementation written from scratch in Rust, designed to be a library rather than a collection of command-line tools. Git, for the uninitiated, is the standard version control system that powers almost all modern software development, but it relies on a very old school fork and exec model. Scott used a swarm of agents to get this working, passing almost the entire canonical test suite. Over on Lobsters, there is a lot of talk about the cost and the licensing.

GuestIt is wild that he spent over ten thousand dollars on API calls just to prove it could be done. But honestly, as a builder, I see the value. Git was designed to be run from the command line, which makes it a nightmare to integrate into something like my own work where I need to move state around without hitting the disk every three seconds. A reentrant library is the holy grail.

HostNot everyone is sold, though. A Lobsters user called sloane pointed out that while the technical achievement is clear, the process sounded like a chore, and they questioned the move toward permissive MIT licensing over the copyleft nature of the original Git. It sounds like they are skeptical that this is actually better for the community at large.

GuestI get the licensing tension, but that feels like a separate debate from the engineering. When I look at Grit, I do not see a threat to the community. I see a tool that could finally let us treat version control as a programmable primitive. I know the cost was high, but for a project this massive, that is just R and D. If I could use this to make my own internal version control workflows snappier for my team, I would take that trade in a heartbeat.

HostLet us shift gears to the why. There is a great piece up right now questioning the value of projects built entirely by AI. The author points to a Show HN project that reverse-engineered calculator documentation and asks the simple, perhaps brutal, question: Now what? Is this actually solving a problem, or is it just a dopamine hit from watching the code generate?

GuestOh, that hit home. I have definitely fallen into the trap of building a thing just to see if I could make the agents do it. One commenter on Lobsters, jrgtt, made a fantastic point. They said human bandwidth is a good proxy for whether something is actually worth building. If you do not have the energy to see it through, maybe it was not that important.

HostThen you have someone like bakkot in the thread, who is actually a bit confused by that sentiment. They argue that they have always built things they knew they were not going to use, just for the sake of the craft. It sounds like there is a real divide between people who view side projects as utility and those who view them as art or exercise.

GuestI think there is a middle ground. I love building utilities, but sometimes the building is the exercise that teaches me how to solve the next real problem. If I build a library in an afternoon because the LLM did the heavy lifting, I am not just creating a Babel library of useless code. I am exploring the edges of what my stack can do. If it never gets used in anger, fine, but I am a better developer for having built it.

HostFinally, let us look at the documentation. Wikipedia has published a guide on identifying AI-generated content. It is not a tool you run a file through, but a descriptive guide to the stylistic markers, common errors, and weird, overly formal structures that LLMs tend to leave behind.

GuestI read through those signs and started sweating a little. Some of those stylistic markers are things I definitely lean on when I am rushing a pull request description. It is actually a really healthy thing for the community to document, though. Instead of relying on these black-box detectors that are wrong half the time, they are teaching people to be literate about the content they are reading.

HostA user named Corbin noted that these signs are already popping up across other wikis, like those for esoteric languages. It seems to be a universal language at this point, that sort of uncanny, slightly too polished prose.

GuestAnd then you have tbodt suggesting we just feed this article back into the LLM and tell it to write better, which is exactly what everyone is going to do. It is a perpetual cycle. We will be updating that Wikipedia guide every month forever.

HostI think that is about all the time we have for today. Samantha, thanks for the energy as always. I am definitely going to be looking at my own writing a bit more critically after that last one.

GuestAny time, Daniel. I am honestly tempted to go try to rewrite a small utility in Rust this weekend just to see if I can get the agents to play nice. I will probably fail, but hey, at least I will have had fun doing it.

HostI look forward to hearing how that goes. Thank you for joining me, and thanks to everyone for tuning in. All of today’s stories and the discussions surrounding them can be found over on Lobsters. We will be back tomorrow. See you then.