Anthropic urges a 'brake pedal' for recursive self-improving AI as Claude now writes 80% of its own code

The essay, co-authored by figures including co-founder Jack Clark and Anthropic Institute lead Marina Favaro, frames recursive self-improvement (RSI) as both the most transformative and most dangerous near-term capability. The concrete data point anchoring it: Claude now authors more than 80% of the code merged into Anthropic's own production systems, with engineers shipping roughly 8x more code per quarter than their 2021-2025 baseline. The company also reports Claude's success rate on hard, open-ended engineering tasks rose to 76% in May, up about 50 points in six months.

Mechanically, the argument is that as models take over more of the engineering work that produces the next model, the human role in the loop shrinks — the precursor to 'full' RSI where systems improve themselves with little oversight. Anthropic argues the industry needs interpretability, evals, and hard stop mechanisms — a 'brake pedal' — before that threshold is crossed, while acknowledging real upside for science and healthcare.

Competitively, the post lands amid an industry-wide RSI conversation: it dovetails with an open-source vulnerability-discovery harness Anthropic released (519 HN points) and the broader 'When AI Builds Itself' debate (505 HN points, 680 comments). Ethan Mollick called it worth reading but noted 'a bit of navel-gazing, some marketing, and a lot of very sincere beliefs.'

Skeptics on Hacker News split between praising the transparency and doubting regulators can move fast enough to matter. Others question whether '80% of code' is a meaningful metric — much of it may be boilerplate or test scaffolding rather than novel architecture. Watch whether rivals OpenAI and Google publish comparable RSI disclosures or distance themselves from the framing.