Everything that humanity has coded to date—every operating system, transfer protocol, open-source library—may only represent 1% of what’s about to be built. We’re living through a technological rupture in which the classical software development lifecycle (SDLC) is collapsing as the centre of gravity shifts towards autonomous sub-agents capable of planning, drafting, rebasing, and testing software.
No other platform is more acutely privy to the technological shift than GitHub. In September, over 40% of pull requests on the code-hosting platform were started by AI agents—a leap from 17% in March, according to GitHub Octoverse 2026 report.
“The SDLC is no longer linear. In fact, the SDLC is collapsing. Most of us jump between these different stages depending on the task. Now, agents are handling the bulk of the implementation and the verification work, and everything feeds into and out of this central learning machine,” Jay Parikh, EVP of CoreAI at Microsoft, said during the keynote at GitHub Universe 2026.
This swarm of agentic activity has also put considerable strain on GitHub’s infrastructure. In July, it recorded eight performance-degrading incidents affecting WebUI, Copilot, APIs, and SSH authentication. This was followed by a nearly eight-hour global outage a month later that was traced to a critical infrastructure failure at a US data centre.
“To do all of this, you need a super strong foundation, an infrastructure that is performant and reliable. Now, there are times this year when we struggled with these things. But a lot of that was due to the growth and scale. That’s not meant to be an excuse, though we will, and we must, do better,” Parikh noted.
To keep its 225 million developers and half a billion public repositories from buckling under the weight of autonomous commits, the Microsoft subsidiary has added more than four million cores and 160 petabytes of internal storage this year.
GitHub is also rebuilding core Git primitives by decoupling compute from storage and minimising coordination. This architecture is designed to allow compute and storage workloads to scale independently, with an initial target of improving scale by 180x.
The Commit Scale
According to the Octoverse report, GitHub recorded 34.3 billion commits and 17.9 billion pushes in a single year, with merged pull requests crossing 1.3 billion.
The scale of agentic adoption can be illustrated in how a team of only three Anthropic engineers ported its core Messages API from the legacy Python infrastructure to Rust—which outperforms it in raw speed, memory safety, and concurrency management—in just one month. Crucially, none of the engineers actually knew the language.
“We actually don’t have a single developer who actually knows Rust… this was actually Claude knowing how to write Rust, and us trusting Claude and then building a lot of systems around it,” Angela Jiang, Head of Product, Claude Platform at Anthropic, revealed in a conversation with AIM.
This is why Git’s 20-year-old architecture needs a shake-up. It was designed for decentralised, serialised commits at a human scale, not for fleets of autonomous agents simultaneously rebasing, testing, and compiling across multi-repo projects simultaneously. Such activity can trigger retry storms.
GitHub is now rearchitecting Git entirely to prevent high-concurrency loops from blocking enterprise pipelines.
“We are completing our full multi-region migration into Azure early next year to ensure that the infrastructure scales elastically with customer demand,” Parikh told AIM.
A Slew of Announcements
Beyond the underlying infrastructure investments, GitHub’s product, commercial, and security announcements aim to bolster its end-to-end software delivery system.
Copilot Projects builds a multi-agent management platform, enabling developers to coordinate sub-agents and track parallel work.
Copilot Automations offers a self-improvement feedback loop in which the engine monitors human intervention and automatically refines its prompt sequencing and tool usage. Meanwhile, Copilot Dreaming automatically opens pull requests to update outdated instruction files when underlying codebases change languages or frameworks.
The headline announcement was Copilot Fusion, a runtime model orchestration engine that evaluates task complexity and routes requests to the model best suited to execute it. These include open-weight models like Moonshot AI’s Kimi K3 and Alibaba Qwen 2.5.
“We do offer certain open-weight models in the model picker today, and we’ve had Qwen in there, and we’ve had others in the past. So, we continue to really listen to what developers want and bring that into the model picker,” GitHub COO Kyle Daigle said.
According to the San Francisco-headquartered company, on the Artificial Analysis Coding Agent Index, Fusion Max scored within two points of Claude Opus 5.5 Max while running 40% faster at roughly half the API cost. The company is offering 30% off Copilot Fusion usage till the end of November to encourage adoption.
It is also pushing Copilot Code Review Max as an autonomous code review tier that GitHub touts as having posted the highest F1 score on ReviewBench for critical vulnerabilities. Meanwhile, AI Scans detects vulnerabilities across codebases and runs PR checks, while reachability analysis prioritises security alerts.
Staying Relevant
Across keynotes, demonstrations, and interviews, executive leadership repeatedly insisted that “the developer is still at the centre”—even as every demo showcased automated systems filing issues, reviewing diffs, and merging code.
“AI does not and cannot replace the human heart in this equation. You must continue to bring the creativity, the judgement, and the taste in this new era. AI doesn’t write the story; you write the story,” asserted Martin Woodward, VP of Developer Relations at GitHub.
At the same time, open-source maintainers are burning out. Famously, Daniel Stenberg, the creator of command-line tool cURL, suspended public vulnerability-reporting bounties after being flooded with hallucinated AI-generated bug reports. Many code maintainers are also actively banning unverified LLM-assisted patches due to the additional verification load they create.
“Opening a PR is way too late in the process to ask: ‘Should I build this? Or is this the right architectural approach?’ We should have had that conversation much earlier, before I spent the tokens building something we might need to now rewrite or throw out altogether,” observed Maggie Appleton, Staff Research Engineer at GitHub.
The company works directly with approximately 2,50,000 maintainers worldwide through the GitHub Secure Open Source Fund, in partnership with organisations like the Open Source Security Foundation and the GitHub Security Lab.
But GitHub asserts that despite competition from startups like Cursor Origin or ex-GitHub CEO Thomas Dohmke’s AI programming startup Entire, the platform can navigate infrastructure challenges arising from AI slop. Its defence lies in enterprise lock-ins and the belief that, regardless of how software development evolves, the world would still need code forge—in one form or another.
“I don’t think the world needs a new GitHub. They might need different tools, that’s great, but what they need is us to evolve for this agentic era, which is what we’re doing,” Daigle affirmed.