
GitHub’s Octoverse 2026 report landed October 28 with a stat that everyone is quoting: AI agents started more than 40% of all pull requests on GitHub in September — 110 million agent-initiated PRs in a single month, up from 17% in March. Code velocity tripled. The bottleneck did not disappear. It moved.
The Numbers Are Extraordinary
The scale in this year’s report is hard to absorb. Developers pushed 17.9 billion times in the past 12 months — more than 2024 and 2025 combined. Commits hit 34.3 billion, nearly triple the prior year. GitHub Actions executed 45.4 billion jobs, up 220%, running at 1,440 per second around the clock. Activity spiked sharply in June 2026, roughly when the current generation of coding agents reached practical usability across the platform.
The agent-specific numbers are just as striking. Repositories with agent configuration files averaged 8x more commits than repos without them. Agent-opened PRs close faster — 90% merged within two weeks versus about 86% for human-only work. They are more complex too: 87% of agent PRs span multiple files compared to 79% for human-authored changes. By any velocity metric, agents are pulling weight.
What the Report Does Not Measure
Here is where it gets interesting. The Octoverse measures merge velocity. It does not report bug rates in agent-generated code. It does not track revert rates. There is no data on production incidents attributable to agent-authored commits. GitHub’s own conclusion — “someone still has to come up with the idea and decide whether it works” — acknowledges the gap without quantifying it.
Qodo’s 2026 State of AI Code Quality report fills some of that picture. It found that 89% of organizations experienced AI-related production incidents. Only 3.7% of engineering leaders believe their current quality processes are sufficient as agents expand. Only 35% of developers say agents reliably follow organizational standards, and only 45% of leaders have traceability connecting AI activity to specific code changes.
Merging faster and shipping quality are different metrics. The Octoverse reports one of them.
The Secret Leak Problem Gets Worse at Scale
One finding buried in the security analysis deserves more attention. GitGuardian’s analysis of Octoverse 2026 found that Claude Code-assisted commits leaked secrets at a 3.2% rate versus a 1.5% baseline — more than double. GitHub’s Push Protection ran 5.8 billion scans this year, but there were 34.3 billion total commits. That gap is not an edge case. And 64% of credentials discovered as exposed were still valid more than four years after initial detection, which means most organizations are not rotating after disclosure.
Forty percent of high-severity secrets found on developer laptops now sit in AI tool directories. This is where citizen developers — non-technical employees now building production apps with natural language instructions — create compounding risk. The Octoverse notes that construction, agriculture, and healthcare repos grew “several times over” since 2022. Construction leads all industries in agent-opened PRs at roughly 9%. These are not teams with established secrets management practices.
What This Means for Your Team
The practical implication is straightforward even if the fix is not: code review is now the critical path. If agents are writing first drafts of nearly half your PRs, the quality of your software is determined by the quality of your review process, not your writing process. That is a different hiring profile, a different skills investment, and a different tooling budget.
- Instrument your review pipeline to distinguish agent-originated PRs from human-originated ones — you cannot improve what you do not measure.
- Run Push Protection on every push, not a fraction of them; the 5.8B/34.3B ratio in this year’s report is not acceptable when AI is accelerating output.
- Require human final approval on every merge, regardless of Copilot review score — agents approving agent code creates sycophancy loops that are a known failure mode.
The Octoverse 2026 report is evidence that coding agents work. What it is not is evidence that the surrounding processes have caught up. The teams that treat review as craft — rather than overhead — will be the ones still shipping clean code when the agent PR percentage hits 60%.

