When Claude writes its own tool, review hits a wall #
October 2026, Anthropic published a striking report: Claude, its flagship AI, writes about 90% of the code for Claude Code, its own coding tool. This raises a fundamental question for software development: how does a team review software when the tool itself is the primary author? The traditional line-by-line human review paradigm hits a wall.
The scale of this challenge is staggering. Anthropic reported its monorepo absorbed more than 30,000 daily changes by September, a massive leap from roughly 2,500 May. This translates to over 20 changes per minute, a pace at which human review simply cannot keep up.
This shift demanded a re-evaluation of quality metrics. June to September, Claude Code pull requests more than doubled. Over the same period, the share of those changes that required reversion—being undone—fell by over 40%. The system achieved higher throughput with fewer rollbacks, suggesting a new era for code quality assurance.
The answer is a review assembly line, not one super-reviewer #
Anthropic’s answer to agent-generated code is a layered assembly line, not a single super-reviewer. A change passes through several automated checks: first, agents verify the work by running tests or inspecting outputs; then, parallel AI reviewers hunt for bugs, with findings posted only if enough agents agree. Each finding requires a fix or a reasoned close.
Next, an automated security review and traditional CI checks (like builds) block unsafe changes. Finally, human approval scales with risk: documentation fixes from trusted authors need no peer approval, ordinary product changes need one person, and high-risk code requires two.
Two accountability rules govern this system: an agent’s approval never substitutes for a human’s, and the person who prompts an agent owns the change. This prevents endless agent-review loops; after three rounds, the reviewer asks the author to rethink their approach. This change alone boosted single-round PR passes from 65% to 78%. Every past incident also becomes a test case, continually improving the AI reviewer.
More checks can mean fewer human bottlenecks #
Reliable systems thrive on layered, repeatable checks, not human inspection of every transaction. Think of error correction codes in memory, TCP checksums on internet packets, or double-entry bookkeeping. These mechanisms catch errors automatically, preventing human bottlenecks and ensuring integrity at scale.
Engineering roles shift dramatically. Senior developers move upstream, defining project scope and architectural patterns. Downstream, they design the crucial tests, constraints, and risk controls that guide agentic systems. Human attention becomes the scarce resource, directed toward strategic decisions machines cannot yet make, like whether to build something at all or identifying high-risk code requiring human oversight.
Anthropic’s October 2026 post on Claude Code review offers a glimpse into this future, detailing how its multi-layered pipeline enables over 30,000 merges per day. For further reading, consult How the Claude Code Team Does Code Review. A word of caution: Anthropic is grading its own homework. Its reported performance figures—a doubling of merged pull requests and a 40% drop in revert rates between June and September 2026—are not independently verified. The company also expends significant resources on model usage and custom infrastructure, like its Scion version control system, which may not be feasible for others.
Enjoying this? Get one like it in your inbox each morning.
one email a day · unsubscribe in two clicks · no third-party tracking
Build the escape hatch before opening the firehose #
Anthropic’s layered pipeline offers a practical threshold: teams should not reduce human review until they can test changes, stage releases, use feature flags, and roll back quickly. This mirrors how reliable systems have always operated, with layered checks catching errors before they propagate.
Teams should turn prior incidents into recurring tests for their AI reviewers. Repeated human review comments—those common findings that consume valuable attention—must become durable rules, automated tests, or new tools. This ensures that machines learn from human experience, rather than merely off it.
Organizational implications are profound. Leaders should be measured by their span of accountability, ensuring that responsibility remains clear even as automation expands. Decisions must map to the right level of human oversight, allowing machines to coordinate routine tasks without making human judgment or ownership disappear.
Ultimately, the goal isn't to eliminate humans, but to elevate their role. People shift from line-by-line inspection to making strategic decisions: what to build, which code is too risky for machines alone, which designs set patterns, and whether evidence holds up. A person remains accountable for every change that ships, whoever typed it.
Frequently Asked Questions #
How does Anthropic review code written by Claude?
Anthropic combines agent-run verification, AI bug reviews, security checks, automated tests, risk-based human approvals, and rollback safeguards.
Does Anthropic let AI approve its own code?
No. Anthropic’s stated rule is that agent approval does not count as human approval, and a person remains accountable for each change.
Why can’t engineers review every AI-generated change?
At very high change volumes, line-by-line review consumes more human attention than teams have. Automated checks handle routine risks while people focus on consequential decisions.
Should every company copy Anthropic’s review process?
Not without safeguards. Teams need reliable tests, staged releases, feature flags, and fast rollback before reducing human review.