GitHub published Octoverse 2026 and opened Universe on October 28, and the two tell one story from opposite ends. By September, more than 40% of pull requests on GitHub started with an agent's first draft, up from 17% in March; to carry the load, GitHub says it has added 4 million CPU cores this year and is moving to a multi-region Azure deployment. Code volume now grows with agent activity, not headcount. The supply of people who can read it still grows one hire at a time.
Output now scales with agents, not engineers
In the twelve months to September 30, GitHub recorded 17.9 billion pushes, 34.3 billion commits, 1.3 billion merged pull requests and 45.4 billion Actions jobs. Against its restated prior year (6.7 billion, 12.9 billion, 538 million and 14.0 billion), that is roughly 2.7 times the pushes and commits, 2.4 times the merged PRs and 3.2 times the CI jobs. Developers went from more than 180 million to more than 225 million, led by the US at 38.3 million and India at 32.0 million. A quarter more people; well over twice the output.
The monthly pace at year's end is steeper still. September to September, commits grew about sixfold and merged PRs about 3.9 times, and September alone brought 110 million agent-started PRs, about six times March's count. Last year's report counted more than 1 million PRs from Copilot's coding agent between May and September 2025.
TypeScript stayed first, Python second, and Rust grew fastest among the top languages, at 134%, 118% and 162%. Those figures are year-over-year growth in monthly participating accounts, attributed by each repository's primary language, so one account can count toward several. On last year's narrower definitions, TypeScript grew 66%.
GitHub pays for agent traffic in cores; customers pay in credits
GitHub felt the curve first. CTO Vlad Fedorov wrote in April that a 10x capacity plan begun in October 2025 had become a 30x design target by February, under the rule "availability first, then capacity, then new features." Its monthly availability reports log 2, 6, 4, 10, 9, 6, 8, 5 and 2 incidents with degraded performance from January to September, 52 in all. The May report attributes traffic growth largely to agentic workflows, and an August outage traced to shared infrastructure that had not kept pace with Actions growth.
Universe is the answer in hardware. Jay Parikh, Microsoft's EVP of CoreAI, listed 4 million CPU cores, 160 PB of high-speed storage and 50 Tb/s of network capacity added so far this year, with the Azure migration due to finish in early 2027. The Register calls the cores and storage roadmap items and frames them as a cure for agent-driven outages; GitHub's post mentions no outages. Its engineers explain why agent load differs in kind: "an agent in a tight loop commits or checkpoints after nearly every action."
Customers meet the curve on the invoice. Copilot moved to usage-based billing on June 1, GitHub arguing that a quick chat and a multi-hour agent session had cost the same. A credit is one US cent; a $19 Business seat includes 1,900 credits a month and a $39 Enterprise seat 3,900, pooled, with paid overage on by default and no automatic fallback to cheaper models. At Universe, GitHub made the seat optional: GitHub Enterprise customers can use every Copilot feature without a separate subscription and pay for the credits they use, through sales now and self-serve later.
Copilot Fusion, now generally available, routes and orchestrates work across models at each model's standard rate, with a limited-time 30% discount. GitHub says Fusion Max came within two points of Claude Code with Opus 5.5 Max on the Artificial Analysis Coding Agent Index at about half the cost per task; Satya Nadella put the saving at up to 77%, per The Next Web. We found no independent listing for either. The seat was a ceiling. The credit is a meter, and it runs on exactly the variable the Octoverse shows compounding.
A merge is not a review
Code is now the cheap input; CI capacity, reviewer attention and judgment are the dear ones. CI absorbs the volume first, at 1,440 Actions jobs a second. Agent PRs are wider, with 87% touching multiple files against 79% for the rest, and they merge more readily than they did: 83% of agent-involved PRs merged within two weeks in February and 90% in September, while the rest held near 86%. That has two readings: agents produce more mergeable work, or reviewers have grown comfortable approving it. The report publishes merge rates, not review depth, reverts or defects, so it cannot tell them apart.
The industry's answer is machines reading machines. Usage of AI-assisted code review grew more than 90%, GitHub added a deeper Max tier to Copilot code review, and CodeRabbit raised $143 million at a $1.5 billion valuation in August, arguing that "trust is now the bottleneck." Among Jellyfish's lead adopters of AI-only review, 9% of PRs merge with no human review, against 0.5% at the median company. Machine review fails in its own ways: on September 28, Copilot code review reported jobs complete while posting no review for some customers, unnoticed for lack of monitoring. A green check is a claim, not a proof.
What stays human is what was always the job: deciding what should merge, owning what did, and building the tests and policies that let people skim the routine and read the dangerous. That is where engineering time goes next.
What "agent-started" does and does not count
GitHub defines an agent-started PR as one whose first draft came from "a coding agent that we can detect," calls the count a floor, and does not name its signals. It measures origin, not authorship: a human rewrite of an agent draft still counts, while agent code opened by a human may not register. Over 40% is a floor on agent-drafted PRs, not a share of merged code that machines wrote.
It is also self-reported telemetry from the company selling Copilot, and the baselines moved. Last year's report put 2025 at nearly 1 billion commits and 518.7 million merged PRs from public activity; this year's restatement, adding private repositories and shifting the window, says 12.9 billion and 538 million. The one large outside dataset leans the same way: Jellyfish, across 1,300-plus companies and 110 million PRs, finds agent involvement in 28% of merged PRs at the median company. Vendor data too, but not GitHub's.
The 30,000-foot read
For most of software's history, output scaled with people and so did review, because the author could explain the change. Output now scales with compute and credits; review still scales with attention. GitHub can buy 4 million cores. Nobody can buy 4 million careful readers.
Parikh expects everything written to date to be "less than 1% of the software we'll write in the next five years." Discount the forecast; keep the direction. What a team controls is narrow and sufficient: which changes demand a human read, how much CI an agent may burn, who owns the credit budget, whether overage stays on. What it does not control is GitHub's uptime, model prices or the slope of the curve. Watch revert and defect rates on agent PRs, which the Octoverse does not publish; whether incidents fall once the Azure migration lands; and self-serve credit pricing. Until then the old rule holds: if it merges, someone answers for it.
Sources: GitHub Blog, Octoverse 2026 (activity totals, restated 2024 and 2025 baselines, developer and country counts, agent-started share, volume, definition and floor, merge and multi-file rates, AI review growth, language growth and methodology); GitHub Blog, Universe 2026, by Jay Parikh (capacity added this year, Azure timing, Enterprise Copilot access billed by credits, code review Max, Parikh's forecast); GitHub Changelog, Copilot Fusion (general availability, per-model billing, 30% discount, Artificial Analysis claim); GitHub Blog, Copilot is moving to usage-based billing (June 1 switch to AI credits, seat prices, rationale); GitHub Docs, usage-based billing (credit price, included credits, pooling, default overage, no model fallback); GitHub Blog, An update on GitHub availability (10x and 30x capacity plans, priorities); GitHub availability reports for January, February, March, April, May, June, July, August and September 2026 (monthly incident counts, May traffic attribution, August Actions incident, September code review incident); GitHub Engineering, Building Git infrastructure for agent-scale development (agent commit behavior); GitHub Blog, Octoverse 2025 (prior-year developers, commits, merged PRs, TypeScript growth, coding agent PRs); The Register, October 28 (roadmap framing of the build-out and its link to outages; secondary); The Register, June 12 (status-history incident counts; secondary); The Next Web, Universe coverage (Nadella's cost claim; secondary); The Next Web, CodeRabbit (funding, valuation, CodeRabbit quote; secondary); Jellyfish, AI Engineering Trends (agent involvement and AI-only review across 1,300-plus companies; vendor data). CAVEATS: every Octoverse, availability and billing figure is company-reported and single-sourced to GitHub. Agent-started counts only agents GitHub can detect, by undisclosed signals, and records who drafted a PR, not who wrote the merged code. Restated 2024 and 2025 baselines differ from last year's published figures, so cross-edition comparisons are approximate. The Fusion benchmark is GitHub's own claim; Nadella's figure is The Next Web's paraphrase of his LinkedIn post and differs from the changelog's figure of about half the cost, and we found no independent Artificial Analysis listing for Fusion. Jellyfish counts merged PRs with any agent involvement, a different definition from GitHub's, and does not publish its cutoffs for leading companies or lead adopters. The 52-incident total sums GitHub's monthly reports, which count incidents with degraded performance; GitHub's status history logs more (26 in April alone, per The Register). Corrections and clarifications to the leads: 538M prior-year merged PRs is GitHub's restatement (last year's report said 518.7M, public only), and the report's tripling claim compares with 2024's 402M, while year-over-year growth is about 2.4 times. The language percentages measure growth in monthly participating accounts from September 2025 to September 2026, attributed by repository primary language; GitHub's prose says Rust roughly doubled while its chart shows 162%. GitHub says the 4M cores, 160 PB and 50 Tb/s were added so far this year; The Register calls the cores and storage roadmap items. The Universe post does not call Fusion generally available; the changelog does, and describes routing plus multi-model orchestration rather than a simple router. The link between the build-out and agent-driven outages is The Register's framing; GitHub's Universe post cites no outages, though its availability reports and CTO post tie capacity strain to agentic growth.