How Claude Fixes It's Own Code

Summary

The video examines how software teams should review code when AI agents generate changes at very high volume. Its central idea is to shift from manual, line-by-line review toward automated verification, risk-based human oversight, and fast recovery when mistakes get through—while retaining human judgment for architecture, ownership, and consequential decisions.

Anthropic’s Code-Review Process

Theo discusses an Anthropic article describing how the Claude Code team adapted its workflow as agent-written changes increased. The article reports that changes landing in Anthropic’s monorepo rose from about 2,500 per day in May to more than 30,000 in September. From June to September, merged PRs to Claude Code’s repository more than doubled, while the share later reverted fell by over 40%.

The process aims to minimize human review time without increasing reverts or incidents. A typical change passes through several layers:

1. Verify before opening a PR: Run tests, builds, and linting; use fresh-context agents to review the diff; and, for UI work, compare screenshots or test the app directly.
2. Automated review: Claude reviews the change against the repository and the PR’s intent. Multiple agents can search for bugs, while others verify and rank findings. Style-only or low-value comments are filtered out.
3. Security review and CI: Security findings and ordinary checks can block a merge until issues are fixed or explicitly resolved.
4. Risk-based human approval: Low-risk changes from trusted contributors may not need human approval; higher-risk changes require one or more people. Sensitive areas may require a person regardless of other checks. Agent approval does not count as human approval.
5. Controlled rollout and recovery: Merge queues, feature flags, staged releases, telemetry, and quick fixes or rollbacks help limit the impact of failures.

The article treats human time and attention as scarce resources. It tracks measures such as approvals per PR, human comments, time waiting for review, reverts, and incidents. It also emphasizes that the author remains responsible for a change, even when an agent wrote the code.

Improving Review Quality and Preventing Endless Loops

Automated review can catch more bugs, but noisy findings may waste time or prompt agents to add unnecessary code. Anthropic’s team tests reviewers against past incidents and clean PRs, and monitors whether findings are fixed, rejected, or deemed unimportant.

The article also describes controls to help review rounds converge: stop repeated findings, focus later reviews on the new diff, cap review cycles, and ask the author to reconsider the approach after several rounds. After making these changes, the team reported:

• A 13% increase in PRs passing after one review-and-fix round.
• A reduction in PRs needing five or more rounds, from 9% to 1.5%.
• A decrease in average review rounds from 2.08 to 1.32.

Incidents and reverts can be turned into evaluation cases so reviewers can be tested and improved over time.

Theo’s Analysis and Recommendations

Theo agrees with moving routine checking to agents, but argues that review is not a substitute for making code easy to work with and verify. He recommends improving the codebase and its architecture first, then adding static checks such as lint rules and type checks. Skills, review files, and agent instructions should come later, when the underlying code and checks cannot address a recurring problem.

Other points from his commentary include:

• Agents should verify their work with meaningful tests—not simply report that they are done. For complex changes, this may mean running the app, testing in a browser or simulator, capturing screenshots, or recording a demonstration.
• Review bots should verify findings before asking humans to act on them. Theo criticizes workflows that leave people to copy suggested fixes from review comments into another agent.
• Risk assessment should describe a change’s impact rather than automatically assigning particular human reviewers.
• Small changes, reversibility, and quick rollback reduce the cost of mistakes. Slowing down to avoid all errors is less useful than detecting and fixing errors quickly.
• Human review remains valuable for deciding whether work should be done, evaluating scope and architecture, checking risky changes and test evidence, and building shared understanding.
• AI-generated explanations or visualizations of a code area can help teammates understand changes more efficiently than reading trivial diffs line by line.
• Teams should be willing to close or reshape PRs when a simpler or better direction emerges; cheap implementation makes abandoned work less costly.

Theo says his T3 Code team has seen as many as 800 PRs filed in a day. He describes using multiple models for review and building workflows that let agents test changes in a running application. He also considers Anthropic’s code-review product potentially expensive for some teams, while acknowledging the value of its layered review approach.

Practical Workflow

The video highlights these practices:

• Give agents concrete ways to run and verify their changes before opening a PR.
• Include test instructions, risk, and rollback information in PRs.
• Use automated bug and security reviews, but filter and validate findings.
• Match review depth and human approval to the risk of the change.
• Cap automated review rounds and measure whether review changes actually improve results.
• Use staged releases, monitoring, and reliable rollback paths.
• Turn incidents into repeatable tests or evaluation cases.

Sponsor Mention

The video also features Browserbase, presented as a way for agents to search the web, retrieve page content in formats such as HTML, JSON, or Markdown, and interact with websites through a browser—including navigating, clicking, and filling forms.

Main Speakers and Sources

• Theo (T3.gg): Host and commentator, comparing the Anthropic process with his team’s practices.
• Anthropic’s Claude Code team and its code-review article: The primary source for the workflow and reported metrics discussed in the video.
• Browserbase: Sponsor segment describing browser tools for AI agents.