In When AI builds itself (May 2026, Marina Favaro and Jack Clark), Anthropic documents that the development of its own systems is increasingly carried out by the systems themselves, and defines recursive self-improvement as “an AI system capable of fully autonomously designing and developing its own successor”. The figures it reports are internal measurements, already in hand.

The numbers

These are figures self-reported by the lab, which is exactly what makes them relevant to governance.

  • As of May 2026, “more than 80% of the code we merge into Anthropic’s codebase was authored by Claude”.
  • The length of tasks models reliably complete on their own “has been doubling roughly every four months, up from an earlier trend of doubling every seven months”.
  • The cited progression: in March 2024 Claude Opus 3 could complete software tasks taking humans about four minutes; a year later Claude Sonnet 3.7 managed tasks of about an hour and a half; a year after that, Claude Opus 4.6 managed 12-hour tasks.
  • On an internal research agent, the speedup over the starting code grew from ~3x (Claude Opus 4, May 2025) to ~52x (Claude Mythos Preview, April 2026).
  • In the second quarter of 2026 the typical engineer was merging eight times as much code per day as in 2024, with a self-estimated productivity uplift of around 4x.
  • On the most open-ended tasks, Claude’s success rate reached 76% in May 2026.
  • In Project Glasswing, in its first weeks, Mythos Preview found “more than ten thousand high- and critical-severity software vulnerabilities”.

The shifting bottleneck

Anthropic sums up the dynamic through Edison: “Edison said that genius is 1% inspiration and 99% perspiration. But we see perspiration becoming increasingly automated.” When execution is automated, the constraint shifts to supervision. The document says it plainly: “Humans play a substantially diminished role in their development, likely moving most of our effort towards oversight, validation, and verification.”

The verifiability of a pause

On governance, Anthropic proposes slowdown as an option: “We believe it would be good for the world to have the option to slow or temporarily pause frontier AI development.” And it ties the commitment to verifiable reciprocity: “we would slow down or temporarily pause, if other developers at or near the frontier also did so in a verifiable manner.” A pause only counts if it is observable from the outside; a commitment no one can verify stays an announcement. Verifiability is the technical precondition for a slowdown.

The technical substrate of verification

At runtime, oversight, validation and verification turn into concrete capabilities. Observability of every action an agent takes, behavioural control to interrupt or limit its execution, a kill switch, model fingerprinting and provenance, append-only forensic evidence. These are the governance controls of Admina (admina.org) and the Secure and Governed criteria of OISG (oisg.ai): measurable detection and containment times, immutable evidence for a supervisory authority.

It is also why I work on DebugABot: infrastructure for debugging autonomous AI agents and embodied robots at runtime — kill switch, behavioural control and model fingerprinting. If verification becomes the main human role, you need something to exercise it with.

For now, the exponential is measured by those who build it: the 8x, the 80%, the 52x are views internal to the lab. The instruments to observe it and stop it from the outside, at runtime and independently, are the part still being built.

Cover image: Romanesco broccoli — photo by Ivar Leidus, CC BY-SA 4.0, via Wikimedia Commons.