Agents Now Start 40% of Pull Requests: The New Developer Job

3 November 20268 min readBrett Iredale

GitHub's Octoverse 2026 says agents start 40% of pull requests, and Dice sees AI job ads up 91%. Why review skills are now the hottest thing in tech hiring.

Two reports landed in the past week that, read side by side, tell the clearest story yet about where software jobs are going. And I think it's a brilliant story for anyone willing to lean in.

The first is GitHub's Octoverse 2026, published on 28 October. The headline that stopped me in my tracks: by September 2026, more than 40% of pull requests on GitHub started with a first draft written by an AI agent. In March it was 17%. That's 110 million agent-started pull requests in September alone, about six times March's volume.

The second is the Dice October 2026 Tech Jobs Report, released this month. US tech job postings rose 4% month on month and 20% year on year. AI and machine learning postings grew 91% year on year, more than four times the overall rate. And among the skills growing more than 200% year on year, employers are now naming specific AI coding tools in their job ads, including Claude AI and GitHub Copilot, right alongside technical debt, maintainability, design reviews and software technical review.

Put those together and the message is simple. Agents write more of the first draft. Employers want people who can direct them and judge what comes back. That's the new developer job, and it's hiring.

What the Octoverse numbers actually show

The scale of activity this year is hard to get your head around. GitHub says developers pushed code 17.9 billion times in the 12 months to 30 September, more than 2024 and 2025 combined. Commits across public and private repositories hit 34.3 billion, more than three times the level two years ago. Merged pull requests more than tripled to 1.3 billion, roughly 3.5 million a day. GitHub now counts more than 225 million developers, with the US on 38.3 million and India on 32 million.

GitHub ties the step change to a moment around June, when agents moved from single-session helpers to persistent tools that track progress across days, orchestrate other specialised agents and run in the cloud as well as on a laptop. Among GitHub Copilot's monthly active users, 55% now use agentic features, up from 35% a year earlier. The most hands-off mode, where you assign a task and the agent works on it independently, grew 230% year on year.

Here's the detail I love: the agent work is landing. In September, 90% of agent-involved pull requests merged within two weeks, compared with about 86% of pull requests without detected agent involvement. And among repositories created this year, the average repo with an agent configuration file logged eight times as many commits as one without.

There's also a quietly important signal about choice. In September, 61% of Copilot users who picked their own model used more than one, and no single model carried more than 25% of requests. Developers are mixing deep reasoning, general purpose and fast lightweight models depending on the step. At GitHub Universe the same week, GitHub launched Copilot Fusion, which picks the model, or mix of models, for each step of a coding task.

Review is the new superpower

The line in the Octoverse report that every developer and hiring manager should pin to their wall is this one: "Review is how developers shape what comes back."

GitHub backs it with numbers. Stacked pull requests, which break big changes into smaller ones that are easier to review in order, launched in preview on 30 July and more than 264,000 developers had used them by September. Use of AI-assisted code review and GitHub Advanced Security both grew more than 90% year on year. Push protection, which screens for exposed secrets like API keys, ran 5.8 billion scans this year, nearly four times 2025's pace.

That matches the Dice data almost perfectly. Dice says the fast-growing skills cluster around reviewing and maintaining AI-assisted code: technical debt, maintainability, design reviews, software technical review and defect management. It also flags telemetry, traceability and exception handling, the skills you need to monitor and troubleshoot systems once they're live, plus AI adoption and enterprise integration, the work of moving AI from pilot into everyday use.

In other words, the market isn't paying for people who type the most lines. It's paying for people with taste. The engineer who can look at a 40-file agent pull request and say "this is right, this is wrong, and this will hurt us in six months" is worth more today than at any point I can remember.

Two engineers reviewing an AI agent's work on a floating holographic panel on a sunny tech campus balcony, with an electric shuttle passing below

Who's hiring, according to Dice

The job titles Dice saw growing more than 150% month on month are telling. AI Trainers, Machine Learning Data Engineers and Machine Learning Operations Engineers support training and running models. Lead, Staff and Distinguished Engineer titles point to demand for senior technical leadership. And Mainframe Systems Programmers, SQL Database Administrators, Workday Managers and Salesforce Business Analysts keep the core business systems running that AI projects plug into.

That last group is a lovely reminder that "AI jobs" aren't only at AI labs. Dice saw insurance and software lead month-on-month industry growth, with tech postings up more than 30%, and healthcare up 15%. Year on year, software, aerospace and manufacturing were all at or above 50% growth. GitHub's industry data rhymes with this: public repositories serving construction, health care, mining and agriculture have each grown several times over since 2022, and construction has the highest share of agent-opened pull requests of any sector, at about 9%.

What I'd do as a developer right now

Make agents part of your daily workflow. If you haven't yet, pick a coding agent and use it on real work this week. Claude Code, OpenAI's Codex, GitHub Copilot and Gemini CLI are all among the fastest-growing repositories on GitHub this year, so there's plenty to choose from, and general assistants like Grok Bot are handy for planning and explaining code too. Try more than one. GitHub's own data says developers increasingly do.

Get seriously good at review. Practise reviewing agent output the way a senior engineer reviews a junior's work: check the intent, the edge cases, the tests, the security and the long-term cost. Learn to split big changes into small, reviewable steps. This is the skill Dice is seeing in job ads, so put concrete examples of it on your CV.

Write great specs. If the first move is now "direct an agent", then clear, well-scoped instructions are a core engineering skill. An agent configuration file, a crisp issue and good acceptance tests will make you faster than any keyboard shortcut.

Learn to run things in production. Telemetry, tracing and exception handling are growing fast in job ads for a reason. Agents can ship code quickly; someone needs to know when it misbehaves and why.

Consider TypeScript, Python or Rust if you're choosing. TypeScript stayed the most used language on GitHub, Python came second, and Rust doubled in a year, the fastest growth among the top languages. These communities also show the highest share of AI-involved merged pull requests, around 40% to 44%.

What I'd do as a hiring manager

Update your job ads and interviews for how work is really done now. If your team uses agents, say so and name the tools, as more employers are, according to Dice. Then test for judgement: give candidates an agent-written pull request and ask them to review it. You'll learn more in 30 minutes than from any whiteboard puzzle.

Invest in the review layer too. More agent-written code means you need senior engineers with time to review, good automated checks, and a culture where smaller, stacked changes are the norm. Dice's surge in Lead, Staff and Distinguished Engineer titles suggests plenty of companies have already worked this out.

And don't forget the people who know your systems. The jump in demand for mainframe programmers, database administrators and business systems analysts shows that AI projects succeed or fail on the plumbing they connect to.

The big picture

Microsoft's Jay Parikh wrote that he thinks "all the software we've written to date will amount to less than 1% of the software we'll write in the next five years," as quoted by The Next Web. Whether or not you take that literally, the direction is obvious. There will be vastly more software, built by more people, in more industries. Every piece of it still needs a human to decide what it should do and whether it's good.

As GitHub put it, someone still has to come up with the idea, and someone still has to decide whether it works, whether it belongs and whether it's good. That someone is you. If you want to find a team that's building this way, have a look at the latest roles on our jobs board.