AI is increasing the amount of code developers write, which makes testing, review, merging, and deployment the new bottleneck.
2
The outer loop needs tools that prioritize pull requests, assist reviewers, optimize CI and merge queues, and improve deployment workflows.
3
AI code review needs high signal, codebase context, CI integration, and human judgment because generated feedback is not perfect.
Summary
Tomas Reimers describes software development as two loops. The inner loop is where developers write and refine code. The outer loop tests, reviews, merges, and deploys it. AI is changing the inner loop by increasing the volume of code developers produce, while leaving the outer loop with the same people and processes. Reimers argues that this will make review and delivery the bottleneck. He proposes an AI-native outer loop with better pull request prioritization, reviewer assistance, CI pipelines, merge queues, and deployment tools. Graphite's Diamond applies this approach to code review by using codebase and change-history context, summarizing and prioritizing changes, and integrating with CI and testing systems. Reimers says AI feedback still needs human involvement because bots are imperfect. He reports that Diamond comments were accepted in about 52% of cases, compared with about 45% to 50% for human comments, while fewer than 4% were downloaded.
Software development has an inner loop and an outer loop
Reimers divides software development into two loops. In the inner loop, developers get code working and shape a feature until it meets their needs. They then move it to the outer loop, where the code is tested, reviewed, merged, and deployed. This distinction explains why faster code generation does not automatically produce faster software delivery. If the inner loop accelerates while the outer loop keeps the same capacity, work piles up after development. The talk uses this model to frame the rest of the discussion about AI-assisted coding and the tools needed around it.
AI is increasing code volume faster than the delivery process can absorb
Reimers says more developers are using AI and producing larger amounts of code. He cites a GitHub developer survey in which nearly every developer surveyed used AI tools inside or outside work, and says 46% of code on GitHub is being written by Copilot. The exact future share may be debated, but he expects the direction to remain the same: more code will be written by AI. That creates work outside the editor. Developers must review, test, merge, and deploy the larger volume, so the outer loop becomes the bottleneck.
AI-generated code can introduce ordinary mistakes and security problems
When Graphite began building Diamond, Reimers says the team saw widespread AI adoption inside its own organization alongside serious risks. AI can hallucinate and make mistakes, but Reimers says security vulnerabilities were even more worrying. Tools such as Cursor, Windsurf, Copilot, v0, and Bolt were producing more code than developers were used to handling. The same developers then had to review, test, merge, and deploy that extra code. His concern is practical: if the delivery process does not change, the volume created in the inner loop can cause the whole workflow to break down.
The new outer loop needs help with prioritization, review, CI, and deployment
Reimers says the pressure created by AI makes problems familiar from large companies affect companies of all sizes. Teams need ways to prioritize, track, and receive notifications about pull requests. Reviewers need driver-assist features that help them focus and make review more efficient. CI pipelines and merge queues need to handle a high volume of changes. Deployment tools also need improvement. These are separate parts of the delivery process, but they all have to cope with the same increase in code volume. His proposed outer loop covers the path from an opened pull request through production.
AI can streamline the manual work around code review
Reimers's AI-first view is that some problems created by AI-generated code may also be reduced with AI. He wants self-driving code review solutions to take over manual and painful parts of review, leaving developers more time to check whether the product reaches users and whether features work as expected. He is careful about the limits of this idea. AI-generated feedback is not perfect, so bots alone are insufficient. Adding AI teammates, background agents, or reviewers may be part of the workflow, but the surrounding toolchain also needs to be designed for AI-assisted development.
Reimers argues that an AI-native development workflow cannot stop at the IDE. If developers become far more productive and generate much more code, the tools for review, testing, merging, and deployment must reflect that volume. Graphite built Diamond as an AI code review platform with what Reimers calls high signal and low noise. It uses an understanding of the codebase and change history to summarize, prioritize, and review each change. It also connects to CI and testing infrastructure so it can summarize errors and help correct failures. The design keeps human attention on the parts of a change that matter most.
Diamond's reported feedback is accepted at a higher rate than human comments
Reimers gives two measures for Diamond's review comments. He says fewer than 4% of the comments are downloaded, and that accepted comments are integrated into pull requests at a higher rate than human comments. He puts human comment acceptance at about 45% to 50%, while Diamond comments were accepted at about 52%. He says the team has spent substantial time tuning that number and that the 52% figure was new as of March. These figures describe Graphite's own experience with its product, rather than a general claim about all AI code review.
"AI can hallucinate. It can make mistakes and almost more scarily it can make security vulnerabilities."01:46
Who should watch
You are increasing AI-assisted code generation and review, testing, or deployment is starting to slow releases.
Your team is deciding how to add AI to pull request review without treating bot feedback as a replacement for engineers.
You maintain CI pipelines or merge queues that were designed for a lower volume of changes.
