GitHub’s executive interview after Universe presents a platform rebuilt at every layer for a flood of agent work. Primary data confirms the flood, then complicates the story: an outage lasting 7 hours 47 minutes, a missed migration target, and a cost benchmark that holds only inside its test setup. Teams relying on GitHub need the conditions attached to each claim.
Table of Contents
AI coding agents now start a large share of the work on the biggest code host in the world. GitHub reported on 28 October 2026 that agents started more than 40% of pull requests in September, up from 17% in March. That is about 110 million agent-started pull requests in one month, roughly six times the March volume. Across the 12 months to 30 September, developers pushed code 17.9 billion times and made 34.3 billion commits, more than three times the commit count of two years earlier, and the platform counts more than 225 million developers.
The fullest account of what that growth demands came in an interview recorded after GitHub’s Universe keynote, with Jay Parikh, the Microsoft executive vice president who oversees GitHub, on the Latent Space channel. It is a conversation with the person responsible for the platform, so its figures are GitHub’s own. This article tests them against GitHub’s published reports, press coverage and an outside benchmark. The adoption numbers hold. The infrastructure and cost claims need conditions.
Adoption figures that survive a check
The totals in the interview match GitHub’s Octoverse post line by line: 17.9 billion pushes, 34.3 billion commits, 1.3 billion merged pull requests and 225 million developers, with the United States at 38.3 million and India at 32.0 million. A new developer every second adds up to 31.5 million a year, although that pace comes from the keynote rather than the written report.
Two details need their units. Parikh says open-source contributions are up two and a half times, but the post gives 2.5 billion this year against 1.12 billion last year, a rise of about 2.2 times. And “six times in six months” describes the count of agent-started pull requests, not their share, which went from 17% to more than 40%.
GitHub counts a pull request as agent-started when a detectable agent wrote the first draft, and its own glossary calls that a floor. Agent pull requests merged within two weeks at 90% in September, against about 86% for the rest, but a merge rate measures acceptance. The post gives no figures on defects or reverts.
Growth that outran the servers
GitHub’s own post-mortems show where the load went. On 17 August an outage lasted 7 hours 47 minutes and disrupted github.com, authentication, Actions, the API, pull requests, issues and Copilot. At its worst about one in five web and API requests failed, and roughly half of archive and raw-content downloads, according to BleepingComputer’s reading of the status page. GitHub’s chief technology officer wrote that neither that outage nor the Actions failure on 6 August came from a code or configuration change: both were capacity failures, and a retry loop in Copilot services added traffic during recovery.
Parikh’s explanation is growth. GitHub says monthly commits rose from 1.4 billion in April to 2.9 billion by August, and its October engineering post then puts September alone at 7.38 billion commits, more than five times a year earlier. Neither post explains how the two figures connect, so the pace of the surge is less settled than the annual totals. The strain also continued after August: on 13 September a background cleanup job saturated a shared database, and GitHub says about 96% of web requests to create issues were affected.
A migration target already missed
Parikh’s answer on timing is short: the move to Azure completes early next year. GitHub’s September availability report shows how much remained at the start of October. By 7 October, 120 of 160 production services were deployed in Central US and 13 of 51 primary MySQL databases had moved, and GitHub states that it missed its 30 September goal of deploying every production service there. It had paused Central US deployments on 11 September because capacity was uneven across availability zones.
Traffic percentages look better but measure different things. Azure handled 65.4% of read traffic for the core application, 48.3% of Git read traffic and 80.2% of stream processing, and GitHub warns against adding them together. In August the chief technology officer put Azure’s share of platform load at roughly 58%, up from 12% in May. TechTimes reported in August a goal of getting all production traffic out of GitHub’s own data centers by the end of 2026; the keynote date is later, though the two may differ in scope.
The Git rebuild carries its own two numbers. GitHub’s engineering post cites up to 35 times higher write throughput in internal benchmarks, while a community recap of the keynote reports a test of 7,200 writes a second from 5,000 agents, described as 180 times current throughput. Both are internal tests, and neither is a guarantee for a given repository.
Model routing saves money under conditions
Parikh presents Copilot Fusion as the fix for a cost problem: choose the right model for each step and developers get quality quickly and more cheaply. GitHub’s claim is that Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code running Opus 5.5 Max, at about half the cost and 40% less time per task. Artificial Analysis, which ran the benchmark, confirms the cost side: $5.59 per task against $11.42, with about half the tokens at 7.2 million per task.
The conditions matter. Artificial Analysis tested Fusion in GitHub Copilot’s command-line tool against Opus in Claude Code, so the comparison mixes routing with two different agent setups. Fusion leads Terminal-Bench 4.0 at 71 but ranks third on FrontierCode at 51, behind 56 for Claude Code with Opus 5.5, and its SWE-Atlas score of 57 lags other systems. Billing follows each underlying model’s standard rate, and a 30% discount on paid plans runs only until 30 November.
Octoverse supports the premise behind the product: in September 61% of Copilot users who pick models used more than one, and no single model took more than 25% of their requests.
The pull request becomes a pressure point
Asked about the argument that agents are ending the pull request, Parikh answers with a platform position rather than a prediction. Some of his own repositories run without pull requests or issues, and GitHub serves 225 million developers who work differently, so these concepts stay configurable. The product record backs him: Octoverse says more than 264,000 developers used stacked pull requests, a preview launched on 30 July, to split large agent output into reviewable pieces.
The pressure falls on maintainers. A community recap of the keynote says repositories can now cap open pull requests from outsiders, exempt established contributors by account age or merge history in an update dated 28 October, and since February restrict or disable pull requests altogether. AIM adds that the creator of cURL suspended public vulnerability bounties after a flood of AI-generated reports.
On dependencies, Parikh rejects replacing 2,000 packages with 20, since the other 1,980 would still need maintaining, and says open source will count for more to businesses, not less. GitHub’s own answer is a package firewall for npm, in public preview, which screens downloads by package age and malware advisories.
Planning for a platform still scaling
For an engineering lead the useful reading is conditional. Git itself has been the steadier layer. GitHub’s 90-day status figures, as TechTimes reported them on 18 August, put Git operations at 99.99% and Actions at 99.33%, roughly 14.5 hours of downtime in three months, and the dashboard counts degraded performance as zero downtime. Teams whose delivery runs entirely through Actions have a reason to keep a fallback pipeline and to cache dependencies instead of pulling them from GitHub at build time.
Spending needs the same discipline. Enterprise customers can now use Copilot features without a separate subscription and pay for the AI credits they consume. Parikh’s test is outcomes rather than tokens: cost savings, engagement, retention or customer satisfaction, because teams “should be attached to these outcome metrics”.
The evidence supports GitHub’s growth numbers and its direction. It does not yet support the timelines. The next availability reports and the Azure completion promised for early next year are the observable tests, and a quarter without a capacity outage would change this reading more than any keynote figure.
Author:
Jan Bielik
CEO & Founder of Webiano Digital & Marketing Agency
This article is an original analysis supported by the sources cited below
GitHub’s Growth: 225M Developers, 34.3B Commits — Jay Parikh, EVP of CoreAI, Microsoft
The interview itself, recorded after the Universe keynote, supplying Parikh’s claims on adoption, Fusion, the Azure timeline, pull requests and dependencies.
Octoverse: Commits, pushes, reach new highs in 2026 as developers build with agents
GitHub’s annual report giving the 17.9 billion pushes, 34.3 billion commits, 17% to 40% agent pull request share, merge rates and stacked pull request usage.
GitHub Universe 2026: Evolving every layer of the stack
The keynote announcement listing Fusion’s benchmark claim, the 4 million cores added this year and the early-next-year Azure completion plan.
GitHub availability report: September 2026
GitHub’s account of the missed 30 September goal, the 120 of 160 services deployed in Central US and the 13 September database incident.
The August 17 outage, and the work ahead
The chief technology officer’s explanation of the 7 hour 47 minute outage as a capacity failure, with commit growth from 1.4 billion to 2.9 billion a month.
Building Git infrastructure for agent-scale development
The engineering post with 7.38 billion commits in September and the internal benchmark of up to 35 times higher write throughput.
GitHub Copilot Fusion orchestrates Anthropic and OpenAI models to balance performance and cost
The independent benchmark behind the cost claim, with $5.59 against $11.42 per task and the weaker FrontierCode and SWE-Atlas results.
Microsoft confirms GitHub is down worldwide
Contemporary reporting of the 17 August outage with the roughly 20% web and API error rate and 50% on archive downloads.
GitHub Actions Hit Three-Nines Failure; One August Outage Consumed Year’s Downtime Budget
The source for the 90-day uptime figures of 99.33% for Actions and 99.99% for Git operations and how GitHub weights degraded incidents.
Does the World Need a New GitHub?
Press coverage quoting Parikh on the early-next-year Azure completion and on open-source maintainer burnout.
GitHub Universe 2026: My recap from San Francisco
A community recap by a GitHub Star, AI-assisted, detailing the maintainer controls, the package firewall and the 7,200 writes a second test.
| Citing this article? Brief excerpts are welcome. Please credit Webiano.digital, name the author where stated, and include a link to https://webiano.digital and to this original article. Full or substantial republication requires prior written permission. Read our Copyright and Content Use Policy. |
This article was prepared with the assistance of artificial intelligence tools. The content underwent expert human review, and Webiano Digital & Marketing Agency assumes editorial responsibility for its final version and publication.
