Anthropic Just Released A New System For Using Claude Code

Highlights

• Anthropic's Claude Code team increased their output to more than twice as many changes while reducing undone changes by over 40%.

• Across Anthropic, the number of daily changes requiring review grew from approximately 2,500 to over 30,000 in a four-month period.

• Giving an agent a set of strict verification tests called evals ensures that work is proven correct rather than merely marked as finished.

• Using a separate reviewer agent with a fresh context window prevents confirmation bias during code reviews.

• Limiting review loops to three rounds and skipping previously checked items prevents agents from getting stuck in endless retry cycles.

Timeline

1. 00:00 - 02:33

   The Review Bottleneck in AI Coding

   Autonomous coding agents can build features for hours without human supervision, creating a massive review bottleneck. When mistakes reach users, developers spend twice as much time fixing bugs as they would have spent reviewing properly. To solve this, the review process must shift from asking if code was read to determining what can go wrong and checking work while the agent builds.

   • AI agents write code quickly but frequently miss requested details.
   • Daily code changes at Anthropic surged from 2,500 to over 30,000 in four months.
   • Rushing reviews shifts debugging time later when fixes become more expensive.
2. 02:33 - 05:43

   Setting Up Evals and Design Mockups

   The first step in verifying agent work is providing a list of tests or evals that bad work will fail. By organizing requirements into specific folders with markdown files, developers can prompt Claude to build eval structures and design mockups. This allows quick visual validation of mobile and desktop layouts before writing actual application code.

   • Evals are fixed, strict tests that prove whether a feature works correctly.
   • Claude can generate feature requirements and eval files automatically using dedicated skills.
   • Creating simple web page mockups before coding prevents token waste and aligns visual designs.
3. 05:43 - 11:44

   Multi-Agent Review Loops and Risk Classification

   Using the same agent that built the code to review it leads to rubber-stamping. Separate reviewer agents read changes using predefined project rules to catch critical flaws. Furthermore, features are classified by risk level—ranging from low-risk visual flaws to high-risk data security issues—so that critical components receive stricter evaluations.

   • A separate reviewer agent with a fresh context window checks code without builder bias.
   • Project rules defined in a review.md file guide automated code review commands.
   • Risk classifiers categorize features by potential customer impact to adjust review strictness.
4. 11:44 - 15:52

   Loop Limits, First-Time User Testing, and Rollbacks

   Unconstrained agent loops can waste tokens by endlessly debating non-issues; enforcing strict iteration caps and focusing only on recent fixes resolves this. To catch remaining mistakes, isolated agents test the live application as new users. Combined with Git branching and targeted debugging skills, teams can quickly isolate failures and roll back problematic updates.

   • Restricting review loops to three or five rounds prevents agents from looping indefinitely over old problems.
   • Fresh sub-agents with blocked code access simulate first-time user experiences to find usability bottlenecks.
   • Git version control and automated bug-fixing skills enable rapid rollbacks when issues emerge.