All of this depends on a foundation that’s fast, reliable, and built to scale. Agents are changing the volume and shape of software development. Adding capacity matters, but capacity alone won’t get us there. We’ve also had to change the architecture.
I’m looking forward to a more reliable GitHub! I think the most important thing I got from the announcement is how they engineer the next GitHub to handle the traffic in the agentic era. Reliability is why Mitchell Hashimoto, the author of Ghostty, gave up on GitHub and moved on in April (Hashimoto 2026):
[…] for the past month I’ve kept a journal where I put an “X” next to every date where a GitHub outage has negatively impacted my ability to work. Almost every day has an X. […] This is no longer a place for serious work if it just blocks you out for hours per day, every day.
Jay Parikh admitted as much in the keynote:
[There] are times this year when we struggled with these things. But a lot of that was due to the growth and scale that you saw earlier. That’s not meant to be an excuse, though. We will and we must do better.
It seems to me they are handling it on two fronts. One is scaling out, by throwing more hardware at it, and moving to a multi-region deployment in Azure, which the keynote calls horizontal scale:
The other is rewriting the software they built the foundation on, what they call the primitives, to make it much more capable: “Three principles guide our new Git architecture: simplify the primitives, decouple storage from compute, and minimize coordination.” Annoyingly, the post doesn’t contain everything they said in the keynote, and what it leaves out are the numbers:
They are not in the two engineering posts it links to either, so I take the keynote as the latest figures. Jay Parikh on stage:
In testing we found 5000 of the angriest coding agents. And we sent them after a mono repo. And this one repo sustained 7200 writes per second.
The 180x is against the old architecture, which topped out at “dozens of pushes a second on a single repository”. And it is coming soon: “Everyone should be migrated to the new architecture by early next [year].”
Compare that with Anthropic, who this month replaced git with their own source control, Scion, for their monorepo (Anthropic 2026):
| GitHub, new Git architecture | Anthropic, Scion | |
|---|---|---|
| Load | 7,200 writes per second, sustained, from 5,000 agents on one monorepo, in testing | over 30,000 merges per day (about one every 3 seconds) on their monorepo, as of September |
| Gain | 180x throughput over the old architecture | 200x smaller footprint (CPU and memory) than their git mirrors |
| For whom | everyone on GitHub, by early next year | Anthropic’s monorepo |
It isn’t really a fair comparison: a merge is not a write, and one gain is throughput while the other is footprint.
Git is not the only thing they are working on. Parikh also went through the AI gateway, which “today processes trillions of tokens”, and GitHub Actions, which “handles billions of minutes every month” and is getting more capacity and simpler coordination. On the network, they are adding more edge sites and working on the backhaul between GitHub’s legacy data centers and Azure.
On Git itself, there was a step before the new architecture. Instead of locking the entire repository for every write, Git now locks “only the branch being updated”, so agents working on different branches move in parallel, which gave 3 to 4 times more throughput. And beyond benchmarks, they took 24 hours of traffic of the busiest open source repository on GitHub, compressed it down to one minute, “and the new stack didn’t break a sweat”.
The harness they build around the Copilot app, the continuity between the desktop and mobile apps, the CLI, VS Code and the web (“Start from your phone at standup, steer it from VS Code at your desk, and check in on it from github.com at lunch”), and the automation pipelines in the cloud are interesting. The last is the most interesting, because it incentivizes people to move some of the work they are already doing in their coding agents (e.g. Codex, Claude Code), like watching GitHub Actions and fixing bugs, into their own native workflow. Copilot automations “run in the cloud, and you can kick them off manually or set them up to respond to an event or run on a schedule”, which is similar to GitHub Actions even if they aren’t naming it that, and each run is billed in GitHub Actions minutes too. The demo chained them: automations that watch for signals and open issues, then triage them, reproduce them, fix them, and verify the fix. And agent merge keeps a pull request going until it can be merged, fixing CI failures and resolving conflicts. Peter Steinberger, the creator of OpenClaw, put the problem on the agents’ side in the keynote:<sup>1</sup>
Many times our agents do silly things like the polling CI […] GitHub [will be] able to tell us when something’s ready.
They also have AI watching for secrets. In the demo, a secret committed on a laptop was caught at git push, before it reached the repository: “this secret was detected by AI. […] I didn’t even have to provide a pattern for it to match.” So it is not hard rules but a supercharged secret scanner, guarding against any kind of secret.
It is nice to see, on a keynote slide,
NOW TOGETHER ON THE SAME LICENSE
GitHub Enterprise & GitHub Copilot
which the post spells out as “GitHub Enterprise customers can use all Copilot features without buying a separate subscription. […] You pay for the AI credits you use.” The same goes elsewhere: Copilot code review “is also available to everyone without a copilot subscription. You just pay for the review with AI credits.” An automation can cap how much it spends, “so I know that my agents aren’t going to go rogue and burn all my AI credits because they’re not free.” And Copilot Fusion, which picks the models for a task, is 30% off until the end of November.
But as an academic, who has GitHub Pro and GitHub Copilot Pro for free (and our UK institutions don’t seem to be paying for subscriptions like this for many of us), today’s announcement is a non-announcement on this front: our access is the same as since the big change in June, when Copilot moved to usage-based billing. With those limits, it is simply not usable: one of my colleagues exhausted his monthly AI credits<sup>2</sup> in a couple of hours on the first day of the month, with the same pattern he used daily before the change.
Quoted from the live captions, which garble him a little.↩︎
GitHub Copilot’s AI credits reset on the 1st of every month.↩︎