Final month, Mitchell Hashimoto, HashiCorp co-founder, publicly declared that he was transferring his fashionable open supply Ghostty terminal emulator undertaking from GitHub. GitHub runs the world’s largest service constructed on the Git distributed model management system, created by Linus Torvalds.
As soon as an enthusiastic person, Hashimoto grew disillusioned with service disruptions, and increasingly slow pull requests. “That is not a spot for critical work if it simply blocks you out for hours per day, every single day,” he wrote.
Hashimoto was fast to defend Git itself: “The difficulty is not Git, it is the infrastructure we depend on round it: points, PRs, Actions, and so forth.”
Many have blamed GitHub’s efficiency on Microsoft, which acquired the corporate in 2018. However to be honest, GitHub itself has been experiencing heavier-than-expected site visitors because of a proliferation of AI-generated pull requests.
In 2025, GitHub saw a 206 percent year-over-year growth in AI-generated tasks measured by way of Bash shell scripts, a widespread approach of working brokers. And extra AI code means extra bugs. Analysis from GitClear found that AI-generated code heaped 10.83 points per pull request, in comparison with 6.45 for the old school human selection.
Our new agentic workforce is elevating huge questions on how the complete software program improvement lifecycle (SDLC) ought to evolve, and if Git ought to come alongside.
“Brokers are nudging us towards a steady stream,” warned Peco Karayanev, co-founder of DevOps platform supplier Autoptic, which bridges Git-based deployments with observability instruments for agent-based remediation.
Autoptic’s total person base runs on some type of Git, both homebrew or from a service supplier like GitLab.
Given the quantity and magnitude of modifications throughout repos, “we want git to start out working in a extra steady mode,” Karayanev wrote in an electronic mail interview.
Git operations, particularly when utilized in GitOps-style automated deployments, nonetheless should be managed by individuals. Updates, commits, pushes, merges are sometimes yoked into sequences of “cease/go” episodes the place somebody has to hit enter on the keyboard just a few instances to proceed the workflow, Karayanev famous. This mannequin could not maintain up as soon as brokers begin getting precedence.
A butler for Git
Git has all the time had its share of critics, particularly those that use the software each day.
There will not be one other piece of software program that’s so broadly adopted and but so inscrutable. Torvalds and different Linux kernel builders constructed Git in 2005 after frustrations with attempting to shoehorn Linux code into the industrial BitKeeper software. Linux, a worldwide group undertaking of mammoth proportions, required a distributed model management system capable of help non-linear improvement of hundreds of parallel branches.
Like several distributed system, Git may be obscure.
One of many co-founders of GitHub, Scott Chacon co-wrote a e-book on utilizing Git (2009’s Pro Git) and nonetheless he finds himself often flummoxed by the model management system.
There are nonetheless “sharp edges” to Git, Chacon informed The Register. “There’s a whole lot of stuff that it does not do very nicely from a usability standpoint,” he mentioned.
Chacon co-founded GitButler as a solution to “rethink the porcelain” of Git, to make Git extra appropriate to trendy workflows. (Final month, GitButler received $17 million in enterprise capital funding).
Consider GitButler as a super-powered Git consumer. It permits the developer to work on two totally different branches concurrently, utilizing a way referred to as digital branching. It reconciles the code a developer is engaged on with the upstream code. They’ll reorder commits, or edit the feedback of a earlier commit. It gives richer metadata concerning the recordsdata being labored on. It will possibly present which commits are distinctive to that department.
Better of all, it eliminates what many builders name “rebase hell,” the place merges into an up to date codebase have to be checked one after the other, an issue GitButler solves by conserving the person’s code synchronized with what’s upstream.
Many of those actions GitButler gives may be carried out by the Git command itself – though Git’s command language, and its guidelines, may be so obtuse that “you’ll in all probability make a mistake sooner or later,” Chacon mentioned.
A Git for brokers
Chacon believes GitHub’s present reliability points stem from the present tsunami of agentic work.
That is “ironic” as a result of GitHub was constructed to scale Git, he mentioned. “However an inflow of brokers is pushing the service to the brink.”
The issue lies not with Git itself, however with everybody utilizing one service, Chacon argued. Final yr, GitHub had about 180 million customers working throughout 630 million repositories – with 121 million created in 2025 alone, in accordance with the corporate’s most up-to-date annual Octoverse report.
“From the longer-term perspective, it does not should be like this,” he argued. Perhaps Git must be run domestically, mirrored globally and managed with purchasers … reminiscent of GitButler, Chacon advised. Maybe Git-based model management methods might be personalized for particular trade verticals.
We want to consider how we “distribute these methods extra,” he mentioned. “Git is designed to be distributed however we’re not distributing it,” he mentioned.
GitButler has created a command line interface particularly for brokers. It was designed to offer MCP servers an built-in map of the repository, which in any other case would require stitching collectively a number of Git instructions. The Digital Recordsdata idea permits the agent to work on a bit of code that can also be being labored on by a developer, or one other agent.
These are modifications that time to a rethinking of how a Git workflow ought to run.
“I believe all of those methods ought to basically change, as a result of all of our workflows have modified, proper? There must be totally different, type of primitives for how one can cope with these drawback units,” Chacon mentioned.
A tip from gaming improvement
One firm that wishes its platform to exchange Git altogether is Diversion, which has constructed an eponymous distributed model management system initially pitched for large-scale recreation design.
“Git’s structure is definitely a difficulty that stops scaling,” argues Diversion CEO Sasha Medvedovsky in an interview with The Register. “Basically it is an structure drawback that may’t be fastened and is a bottleneck for finish customers and internet hosting companies.”
Git is a distributed system insofar as each person, or hosted service, requires a devoted database (very similar to blockchain). “It is not distributed within the common sense however moderately replicated,” he wrote in an trade with The Register on LinkedIn.
Operations run on a single thread, making concurrent operations inconceivable. Because of this, the bigger the repository, the slower the commit operations – a lethal mixture for fast-paced agentic software program improvement, Medvedovsky famous.
After all, each CEO could have their speaking factors prepared a couple of competitor’s weaknesses (Diversion is finalizing a blog post with exhausting numbers about Git and GitHub efficiency). However there are a rising variety of different initiatives round prepping Git for the difficult instances forward.
Maybe most notable is Jujutsu, a Git-compatible distributed model management system, stewarded by Google senior software program engineer Martin von Zweigbergk. Like GitButler, Jujutsu (jj) goals to eradicate a whole lot of the annoyances that include Git. It consists of an undo button and the power to maintain committing even when there’s a battle.
And since every part written in C have to be recast into Rust as of late, long-time Git contributor Sebastian Thiel began a undertaking referred to as Gitoxide to rebuild Git in Rust. Potential advantages embody important performance improvements by multicore processing, and the much-needed memory safety that comes with Rust.
Will Git 3 clear up all the issues?
Git’s chief maintainer is Junio Hamano, who took the reins from Torvalds in 2005. And he stays busy conserving Git present.
At FOSDEM this February, core Git contributor and GitLab engineering supervisor Patrick Steinhardt discussed a number of the modifications coming within the subsequent model of Git, model 3, which is regularly being rolled out this yr.
One of many chief enhancements shall be in the best way Git manages the commit references, the IDs that time to every change being made. Surprisingly, this operation is an actual bottleneck for the software program. “The design is inefficient,” Steinhardt informed the viewers.
Each time a programmer commits a code change, it will get recorded in a “packed-refs” file, which saves time by not giving every commit its personal reference file.
As tasks develop bigger, nonetheless, it takes longer for Git to amend or to delete a reference in packed-refs (One GitLab repo has a packed-refs file of greater than 20 million references, Steinhardt mentioned).
That is particularly problematic when you may have a number of, simultaneous readers and writers of that file. And simply overlook about getting a constant view of all of the references.
The freshly applied Reftable characteristic, which would be the default in Git 3.0, shops references in an indexable binary format. The Git people borrowed this idea from the Eclipse Basis’s JGit Java implementation of Git.
Reftable permits for block updates, eliminating the necessity to rewrite a 2 GB-sized file for a single entry. And it’s a lot sooner for studying, which might pave the best way for Git supporting bigger, extra sprawling repositories – excellent for an ever-busy agentic workforce.
For practically twenty years, Git has proved to be the model management system of selection for geeks worldwide. However even with these new options and numerous third-party enhancements, can it retain relevance for a brand new technology of agentically enhanced coders?
The battle is on. ®