1.2 Billion Commits in a Week, but PR Review Time Up 441%
GitHub received 1.2 billion commits in a week, up more than 400% year over year. But Faros AI data shows that after teams increased AI adoption, PRs spent 441% longer in review. Code generation is no longer the bottleneck; GitHub’s solution is to deploy agents across the four stages of Plan, Build, Verify, and Observe—and along the way, it surfaces a few unanswered questions from the comments.
In October 2026, GitHub received more than 1.2 billion commits in a single week, up more than 400% from a year earlier, and also more than the total number of commits GitHub received in all of 2025.
The numbers are here, and so is the question: do you feel 4x more productive than you were last year?
A Faros AI study offers another angle. After teams went from low AI adoption to high adoption, PRs spent 441% longer in review. Code generation is absurdly fast; the review queue turns into a parking lot. One commenter replied under GitHub’s tweet: code generation was never the slow part.
GitHub recently published a long article explaining its understanding of the agentic engineering workflow, splitting development into four stages: Plan, Build, Verify, and Observe. The original link is at the end. Here are a few points worth reading.
Plan: First, get everyone on the same context
Consider an API change scenario: the design mockup is in one tool, compatibility requirements are buried in a Slack message, and to understand why it is implemented this way now, you have to dig through an old PR. Before you even start writing the plan, just gathering all this information can take hours.
Copilot’s approach is to pull scattered context into one place. Teams can review plans together with product, design, and analysts right in Slack or Teams, and connect external tools to verify assumptions. Once agreed, it generates an Issue with acceptance criteria and breaks it into tasks that can be advanced independently. GitHub Agentic Workflows labels Issues and estimates complexity; low-complexity ones go directly to Copilot, while people handle the parts that require difficult judgment. Complex changes go through Plan Mode, where several approaches are laid out and compared first, and only after approval are they handed to Copilot for implementation.
Build: Don’t Be Tied to a Single Computer
The implementation phase is the easiest to interrupt. The agent is halfway through, a meeting starts, the dog is staring at you by the door, and then another edge case comes up in the meeting that needs handling.
A few design choices are worth noting. Session continuity lets the same Copilot session continue across devices. Tasks run in cloud sandboxes; close the laptop and they keep running. The Copilot app gives each session an isolated copy of the repository, so implementation tasks and documentation updates do not share a half-finished working tree. Repository instructions remember the team’s conventions and validation commands; shared memory stores facts reused across repositories; Dreaming reviews sessions and PR feedback and proposes suggestions for updating memory and instructions. Model selection is handled by Copilot Fusion, which adjusts reasoning depth by task to balance quality, cost, and response time. Enterprise policies follow people through managed settings, so switching clients does not require reconfiguration.
Verify: The “Small Refactor” With 47 Files
Someone opens a PR labeled “small refactor” that changes 47 files. What they want people to look at is the shared validation logic, but that part is buried under changes to dozens of endpoints, formatting diffs, and hundreds of lines of generated tests.
This is what 441% looks like. GitHub compresses from both ends: make changes easier to review, and let machines clear routine checks first.
Stacked pull requests split dependent work into small pieces, each reviewed separately while dependencies remain visible. Code Quality first checks against organizational standards. Copilot code review adjusts analysis depth based on change complexity. AI Scan covers languages and files beyond CodeQL, and AI secret detection finds unstructured secrets such as passwords. License compliance checks run before merge; package firewall, currently in public beta for npm, can block downloads that violate security policies. Agentic Autofix suggests fixes, Agent Merge handles failed checks and conflicts, and once configured, Copilot can merge directly after validation steps pass.
Observe: After All Checks Are Green
The day after release, intermittent timeouts appear on the on-call dashboard, and support says several major customers are affected. You are cross-checking logs, release records, and customer feedback while someone asks whether to roll back. “All checks passed” is not very useful at this point.
Copilot can connect to observability tools and business data such as Fabric through connectors, helping you determine which requests and customer segments are affected. Once the issue is clear, have it draft an Issue with evidence and acceptance criteria, and include regression tests. Dreaming extracts improvement suggestions from sessions and repository activity and feeds them back into memory, skills, and code review guidelines.
The Questions in the Comments That the Article Doesn’t Answer
A few replies under the tweet are more down-to-earth than the article.
Someone asks how safe the data aggregated by models really is. The 1Password development team adds that agentic workflows need to be clear about secrets: agents in GitHub Actions can get keys via OIDC without storing long-lived tokens, which can eliminate one of the most common types of long-lived credentials in pipelines. Someone else mentions that a successful process should be saved as a how-to and called directly next time, rather than starting from scratch every time.
One reply is more direct: has PAC proxy support in Copilot App and Copilot CLI been fixed? The developer’s point is that the avatar is cute, but not being able to use Copilot inside a VPN—and taking away the only method that worked—matters more to thousands of developers.
Start With the Most Annoying Thing This Week
The takeaway of GitHub’s article is pragmatic: don’t replace the entire process at once. Pick something already on your plate, first think about what context it needs, how it should be reviewed and tested, and how to confirm it works after launch—then put an agent into it. Whichever step annoys you most this week, start there.
1.2 billion commits is capacity; 441% is accumulated debt. Both being true at the same time is the real state of development right now.
Original article: https://x.com/i/article/2120953234189946880
发布时间: 2026-11-13 08:29