GitHub published Octoverse 2026 on 28 October. Among all its figures there is one that changes the conversation in any development team: in September 2026, more than 40% of pull requests on the platform started with a first draft written by an agent. In March it was 17%.
The idea this insight argues: if an agent now writes the first draft, what sets a team apart is not writing code. It is what comes before, what is asked for and with which acceptance criteria, and what comes after, what a machine checks and where a person decides because there is risk. The report itself says so in its last section: someone still has to decide "whether it works, whether it belongs, and whether it's good".
What Octoverse 2026 says about agents
Octoverse counts GitHub activity between 1 October 2025 and 30 September 2026, in public and private repositories. These are the figures that touch on how a team works:
Limit These are GitHub's data about its own platform, which includes open source and personal projects: they are not a sample of company teams. GitHub also notes that its agent signal is a floor, because it does not detect all use. And the report gives no figures on defects or failures in agent-written code.
The tools are moving the same way. On the same day, GitHub presented its Copilot app as a tool "designed from the ground up for agentic software delivery", with a tagline that sums it up: "From issue to merge, in one app". This is not a review of the product; it is a sign that the agent is no longer an assistant suggesting lines, but the one opening the change.
Merged more often does not mean more useful
The figure that will be quoted most is the merge rate: agent pull requests are now accepted more often than the rest. It says those drafts are accepted more often, and nothing more. A merged pull request says someone accepted it, not that the change was the one that was needed.
A study of an agent rollout at a large company puts it in these words. Emerson Murphy-Hill, Jenna Butler and Alexandra Savelieva analysed tens of thousands of engineers during the rollout of Claude Code and GitHub Copilot CLI at Microsoft in early 2026: adopters merged roughly 24% more pull requests. And the authors themselves warn that "a merged PR is not the same as the value it delivers". We cover it in more detail in coding agents spread through peers, not by decree.
More drafts accepted with the same number of people reviewing also has a cost: the reviewer's attention. It is the bottleneck we describe in more code, same delivery. DORA sums it up in its 2025 report: AI is "an amplifier, magnifying an organization's existing strengths and weaknesses". A team that already knew what to ask for and how to check it gains; one that did not produces more of the same.
What your team still decides
If the draft is no longer the work, the work moves to both ends of the change. This is how it splits in a team that works with agents:
| Before the draft | The draft | After the draft | |
|---|---|---|---|
| What is done | Decide what changes, why, within which limits and how you will know it is right | Write the code for the change | Check that it works, that it belongs and that it is good |
| Who does it | People in the team, in writing | The agent, more and more (over 40% of pull requests on GitHub) | A machine, for what can be checked; a person, where there is risk |
| What it leaves | A specification with acceptance criteria | A pull request | A recorded decision: merge, fix or discard |
onext, based on Octoverse 2026 (GitHub, 2026-10-28)
The left-hand column is Spec-Driven Development (SDD): writing down what the change has to do before anyone, person or agent, writes the code. It is not new bureaucracy; it is what the agent needs so it does not invent what is missing. The acceptance criteria in that specification are also what a machine can check afterwards, as we explain in your spec is already an eval.
The right-hand column has a clear limit: a machine does not decide everything. Where a person signs off, and why there and not everywhere, is in specs and agents: where you sign off.
Octoverse gives a hint about the teams that use agents most. Among repositories created this year, those with an agent configuration file logged on average eight times as many commits as those without one. It is a correlation and does not prove the file causes anything, but it points in a reasonable direction: whoever works a lot with agents ends up writing down how they want to work with them. What should go in that file, and what should not, is in CLAUDE.md in teams: context, not control.
The agent writes more and more of the first draft. What your team asks for before and what it checks after is still people's work.
Three questions for your team this week
You do not need your team's own figure to start. Look at the last pull requests an agent opened and answer:
- Was anything written down before the draft? If the agent worked from a one-line issue, whatever was missing was decided by the agent.
- What did a machine check and what did a person check? If the answer is "a person looked at everything", review will be the next bottleneck.
- How do you know whether you deliver better? The number of merged pull requests does not tell you. Time to production, failures after deployment and rework do: what to measure is in KPIs for AI development teams.
If you want to answer them calmly with someone who sees it from the outside, the diagnostic is free. Before that, you can read how we install the method in AI-Accelerated Development.
Frequently asked questions
What is Octoverse 2026?
It is GitHub's annual report on activity on its platform, published on 28 October 2026. It covers the year from 1 October 2025 to 30 September 2026 and includes public and private repositories unless stated otherwise.
How many pull requests now start with an agent?
According to Octoverse 2026, in September 2026 more than 40% of pull requests on GitHub started with an agent's first draft, up from 17% in March: around 110 million in September alone. GitHub notes this is a floor, because it does not detect all agent use.
Are agent pull requests merged more often?
In September, 90% of agent pull requests were merged within two weeks, compared with about 86% of those without detected agent involvement. In February it was the other way round: 83% against roughly 86%. A higher merge rate says nothing about whether the change was the right one or how much value it delivers.
Do these figures apply to my company?
As a market signal, yes; as an exact benchmark, no. They are GitHub's data about its own platform, which includes open source and personal projects, so they do not represent any particular company team. What is useful is measuring how delivery changes in your own team, not comparing yourself with the GitHub average.
What has to change in my team if the agent writes the first draft?
What happens before and after the code. Before: what is asked for, with which acceptance criteria and which limits, in other words the specification. After: what a machine checks and at which points a person decides because there is risk.
Does an agent configuration file show up in repository activity?
Octoverse says that, among repositories created this year, those with an agent configuration file logged on average eight times as many commits as those without one. It is a correlation: it does not prove the file causes anything. It does suggest that the teams that use agents most write down early how they want to work with them.
Sources
- GitHub Staff, "Octoverse: Commits, pushes, reach new highs in 2026 as developers build with agents", GitHub Blog, 28 October 2026.
- Jay Parikh, "GitHub Universe 2026: Evolving every layer of the stack", GitHub Blog, 28 October 2026.
- Emerson Murphy-Hill, Jenna Butler and Alexandra Savelieva, "Adoption and Impact of Command-Line AI Coding Agents: A Study of Microsoft's Early 2026 Rollout of Claude Code and GitHub Copilot CLI", arXiv:2607.01418, 1 July 2026.
- DORA, "State of AI-assisted Software Development 2025", 2025.
Bernat López is founder and CEO of onext, an AI boutique. He helps development and product teams work with AI with method —specification before building, a person who decides where there is risk and Spec-Driven Development— and applies to his own company what he proposes: onext runs on its own agentic system.
LinkedIn →