40% of pull requests now start with an AI agent: what it means for career changers in dev
GitHub's Octoverse 2026: over 40% of pull requests now start as an AI agent's draft. The 4 skills that matter most if you're switching careers into dev.
Note. No affiliate links here. Just the numbers from GitHub's Octoverse 2026 report, and what I take from them as a developer who switched careers. More practical resources are in my Tech section. Sources at the bottom.
The number that stopped me
On October 28, GitHub published its annual Octoverse report. One number stopped me in my tracks: in September 2026, more than 40% of pull requests on GitHub started as a draft written by an AI agent. In March, it was 17%.
That's roughly 110 million pull requests in a single month. And GitHub points out that this is a floor, not a ceiling: the platform doesn't detect every time an agent is involved.
When you learned to code past 40, the question is bound to come up: is there still room for us? My answer is yes, just not to do the same job as before.
I already asked that question back in June, in Becoming a developer after 40 when AI already codes. Octoverse 2026 now puts numbers on it, and they point to four skills.
1. Reviewing becomes the heart of the job
GitHub's message is clear: review is where developers shape what comes back. AI-assisted code review grew by more than 90% in a year.
Another telling number: in September, 90% of agent-started pull requests were merged within two weeks, versus about 86% for the rest. So agents write code that gets through. But someone still has to say yes. That someone is you.
It's already how I work with Claude Code: it proposes, I review every change, and I'm the one who approves.
π What you can do: practice reviewing code, your own and other people's. On an open-source project, a well-argued review is often worth more than one more commit.
2. TypeScript is worth even more now
TypeScript is the number one language on GitHub. Among repositories created in 2026 where an agent was involved, TypeScript is the main language in 35% of cases, versus 24% of repositories without an agent.
GitHub's hypothesis: people pick languages that make it easier to check what the agent wrote.
π What you can do: if you're coming from JavaScript, make the move to TypeScript. Types help you catch mistakes, whether the code comes from you or from an agent.
3. Security is no longer optional
GitHub's tool that blocks passwords and access keys from being pushed into code (push protection) ran 5.8 billion checks this year, almost four times as many as in 2025. The more code gets generated, the more chances something leaks.
GitHub's own breach in May drove the point home: one booby-trapped VS Code extension was enough to steal about 3,800 internal repositories. I turned it into 7 security habits for solo developers.
π What you can do: learn the basics of application security. That's exactly why I'm preparing for the CompTIA Security+ certification. And if you haven't yet, start by securing your dev accounts: MFA, SSH keys, clean tokens.
4. Picking the right AI model becomes a skill
In September, 61% of Copilot users who choose their own model were using more than one.
GitHub has just launched Copilot Fusion, which automatically picks one or several models for each task. It's available on every Copilot plan, and each model it uses is billed at its usual rate.
This doesn't contradict my choice to use only a handful of AI agents, but use them well. Comparing several models inside the same tool isn't collecting subscriptions: it's learning to choose.
π What you can do: run the same task through several models and write down what each one does well. Knowing which model to use for what is a skill you can learn.
The bottom line
Agents write more and more of the first draft. Your job is to frame the task, review, secure and decide.
My own summary, based on GitHub's Octoverse 2026 report.
These are exactly the skills where life experience counts. Our career change isn't coming too late: it's arriving right when judgment matters more than typing speed.
Frequently asked questions
What does "agent-started pull request" mean?
It's a pull request whose first draft was written by an AI agent before a human reviewed and merged it. It still counts as agent-started even if someone heavily edits it afterwards. GitHub says its 40% figure is a minimum, because not every agent contribution is detected.
Do I still need to learn to code if agents write the first draft?
Yes. To review a pull request and say yes (or no), you need to understand what the code does. The fundamentals are no longer about writing everything yourself: they're about framing, checking and deciding.
Why learn TypeScript rather than JavaScript in 2026?
Because types make code easier to check, whether it comes from you or from an agent. According to Octoverse 2026, TypeScript is the main language in 35% of new repositories where an agent was involved, versus 24% of repositories without one.
What is Copilot Fusion?
It's a GitHub Copilot mode launched at the end of October 2026 that automatically picks one or several AI models for each task. It's available on every Copilot plan, and each model it uses is billed at its usual rate.
Sources
- GitHub β Octoverse 2026 (annual report, October 28, 2026)
- GitHub Changelog β Route and orchestrate models with GitHub Copilot Fusion (October 28, 2026)
Written on November 6, 2026. These figures come from GitHub's platform data and move fast: check the sources before making any big decision.
What about you: do you actually review the code an AI suggests, or approve it with your eyes closed? You'll find more guides and field reports in the Tech section.