Anthropic scales code reviews using AI agents and risk-based human approval
Anthropic is using a multi-layered AI review system to handle the surge of AI-generated code, reducing manual work while maintaining safety.

The short version
- Merged pull requests to Claude Code more than doubled between June and September.
- The proportion of changes requiring a revert fell by over 40% during this period.
- Low-risk changes like documentation may require no human approval at all.
- Review cycles are capped at three rounds to prevent agents from generating excessive feedback.
- Internal pull requests passing after one review round increased from 65% to 78%.
Why it matters
As AI agents write more software, companies need new ways to verify code without slowing down. Anthropic's model shows how to balance automated speed with human accountability.
Managing the AI code surge
Anthropic recently shared details on how its team manages code reviews as AI agents contribute more frequently to its codebase. The team observed a significant increase in activity, noting that merged pull requests to the Claude Code repository more than doubled between June and September. To handle this volume, they shifted away from traditional line-by-line human reviews toward a system of automated verification.
The goal of this new process is to reduce the time humans spend on manual reviews without increasing production incidents or regressions. A regression is a software bug that causes a previously functioning feature to break. By moving verification earlier in the development cycle, the team aims to catch errors before they reach the final stages of the pipeline.
The team reported that this shift has already yielded measurable results. During the period of increased activity, the proportion of changes that required a revert fell by over 40%. This suggests that the automated layers are successfully filtering out low-quality code before it becomes a problem for the engineering team.
How the AI agents work
The review process begins before a developer even opens a pull request. Claude runs tests, builds, and linters, which are automated tools that check for syntax errors and basic functionality. A separate AI agent then reviews these changes in a fresh context to provide an initial layer of scrutiny.
Once a pull request is opened, it undergoes several automated checks. Claude-powered reviewers inspect the code for potential bugs, and multiple agents independently investigate their findings. Another group of agents then verifies those findings to ensure accuracy. A separate security review can also block a merge if any unresolved issues remain in the code.
To ensure the process remains efficient, Anthropic introduced mechanisms to prevent the AI from generating too much feedback. They capped review cycles at three rounds, after which the author is asked to reconsider the implementation. This prevents the team from getting stuck in infinite loops of automated comments on the same piece of code.
Risk-based human oversight
Human involvement in the process is now determined by the risk associated with a specific change. Low-risk modifications, such as updating documentation or internal features hidden behind feature flags, may require no human approval at all. This allows developers to move quickly on minor tasks without waiting for a manual sign-off.
Changes with medium risk require approval from one person, while high-risk modifications need two separate approvals. Modifications to particularly sensitive areas, such as CI workflows, always require human oversight regardless of the perceived risk. This tiered system ensures that human attention is focused where it matters most for the stability of the product.
Despite the heavy lifting done by AI, human reviewers remain responsible for the final outcome. They still make the ultimate decisions regarding architectural choices and product direction. Anthropic emphasizes that the person directing an AI agent remains accountable for any code the agent produces during the process.
Improving review efficiency
The new workflow has significantly improved the speed of the development cycle. Anthropic reported that the share of internal pull requests passing after just one review and fix round increased from 65% to 78%. This indicates that the AI-assisted process is helping developers get their code correct on the first try.
Conversely, the proportion of requests requiring five or more rounds of feedback dropped from 8.9% to 1.5%. By focusing on the risk of a change rather than the total amount of code a human has to read, the team is optimizing for speed. This shift reflects a broader trend in how engineering teams are rethinking the balance between automation and human oversight.
To keep the system improving, the team incorporates incidents and reverted changes into a growing evaluation dataset. This data is used to train and measure the performance of the automated reviewers. By learning from past mistakes, the AI agents become better at identifying bugs and preventing future regressions in the codebase.
The future of QA
The rise of AI-powered code reviews raises fundamental questions about the boundaries of software engineering. Some observers have asked if code review is becoming synonymous with testing and if the responsibility should move to a dedicated quality assurance (QA) team. These questions highlight the blurring lines between writing code, testing it, and ensuring its overall quality.
Anthropic's approach suggests a model where automated verification handles the bulk of the
What to watch
- Whether other major tech companies adopt similar risk-based human approval models.
- How the definition of a 'QA team' changes as AI agents take over more testing tasks.
Try it
Look for 'feature flags' in your own software projects to see how teams hide new features from the public while they are still being tested.
Sources and links
- Officialclaude.dev/blog/solving-code-review-at-anthropic/ claude.dev
- Postput it x.com
- Reportinginfoq.com: the article we sourced
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by infoq.com · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
