Database › Anthropic
Anthropic
Anthropic published internal research on its own engineers' use of Claude. No mandate is described.[1]
Tools
Metrics tracked
- Anthropic surveyed 132 engineers and researchers, interviewed 53, and analyzed 200,000 internal Claude Code transcripts (August 2025).[1]
- As of May 2026, more than 80% of code merged into Anthropic's codebase was attributed to Claude, up from low single digits before February 2025. Leadership estimates 90%+ including scripts.[2]
- The typical engineer merged about 8x as much code per day in Q2 2026 as in 2024. Anthropic says lines of code overstate the true gain.[2]
Guardrails
- Most respondents said they can 'fully delegate' only 0-20% of their work. Work generally involves active supervision and validation.[1]
- Every proposed change is read by an automated Claude reviewer before merge. A retrospective found it would have caught about a third of the bugs behind past claude.ai incidents.[2]
- Lesse said agents already fix tests autonomously with no human interaction, but leaders are moving carefully on letting agents change the actual codebase on their own.[3]
Hiring, reviews & process changes
- Anthropic says human code review has become a bottleneck as code volume grows.[2]
- Business Insider reported that an Anthropic report from October 2026 said it has hired for a smaller share of junior technical positions as agents took on more work.[3]
Reported results
- Respondents self-reported using Claude in 59% of their work, for a +50% productivity gain. Anthropic saw a 67% increase in merged PRs per engineer per day after adopting Claude Code.[1]
- Claude Code went from about 9.8 consecutive tool calls per transcript to 21.2, and human turns fell from 6.2 to 4.1.[1]
- Employees raised concerns about skill atrophy and reduced mentorship.[1]
- Head of platform engineering Katelyn Lesse said in October 2026 that Claude writes 'literally all' of her team's code, while people still design systems and verifications.[3]
- Superseded (2025 survey): staff now say Claude-written code went from worse than human code in late 2025 to roughly at parity by mid-2026.[2]
Metric tags: Self-reported productivity, PR throughput, Agent autonomy, AI-authored code, % code AI-generated, Code review, Quality / defects
Editor's note: Vendor-authored. Anthropic studies its own product. The 2025 survey figures are superseded by the June 2026 report.
Sources
- Anthropic: How AI is transforming work at Anthropic · 2025-12-02 · vendor source
- Anthropic Institute: When AI builds itself · 2026-06-04 · vendor source
- Business Insider: Top Anthropic engineer says Claude now writes 'literally all' her team's code · 2026-10-22 · press source
Last verified 2026-10-25. Industry and size are editorial bands; see methodology. Spot an error? Submit a correction.