RSI Is the New AGI
This month, Anthropic published internal data that reignited debate over recursive self improvement, the idea that AI systems could eventually design and refine their own successors. The report and the reaction to it were covered by The Batch, the newsletter published by DeepLearning.AI.
The headline numbers are specific to Anthropic. The operational lesson is not.
The Data Anthropic Published
According to Anthropic, 80 percent of the code merged into its systems is now authored by Claude, up from under 5 percent before the research preview of Claude Code. The company reported that by the second quarter of 2026, each engineer was contributing eight times more code per quarter than engineers did in the first quarter of 2023.
In April 2026 alone, Anthropic said its systems shipped more than 800 fixes to its API, cutting API errors by a factor of one thousand, in work the company estimated would have taken a team of humans four years to complete without AI assistance.
Anthropic also tracked how well Claude Code solved problems of increasing difficulty, sorted into four categories: trivial, routine, substantial, and problems where even the solution and the standard for success were unclear. Between September 2025 and May 2026, success rates rose across every category, most sharply on the hardest and least defined problems, where performance climbed from under 20 percent to 76 percent.
These are figures reported by Anthropic, not independently verified by LiveDoc Solutions. They are cited here as reported.
The Debate Anthropic Did Not Settle
Anthropic framed three possible futures. In one, AI stays less capable than top human engineers. In a second, which the company called most likely, AI assisted engineering keeps accelerating while humans retain control of research and development. In a third, AI becomes able to improve itself without that oversight.
Reaction split along familiar lines. OpenAI said it saw early signs of the same pattern in its own systems. Sakana AI, a Japanese research organization, launched a research group called RSI Lab dedicated to the same pursuit. Others were more skeptical. UCLA adjunct professor Arun Rao expects the timeline to be longer than Anthropic projects. AI policy researcher Miles Brundage and MechanizeWork co founder Matthew Barnett both pointed to data and compute limits as a near term ceiling. Wharton professor Ethan Mollick and technology analyst Michael Spencer both noted a marketing dimension to how the report was framed, and connected it to a recent wave of funding into AI startups built around the same narrative.
This section is commentary and forecasting, not established fact. Analysts disagree on how far this trend goes and how fast.
The Part That Applies to Every Business Today
Set the long range debate aside. The measurable part of Anthropic’s report, output per engineer rising sharply because AI is doing more of the production work, is not unique to Anthropic and is not theoretical. It is already underway anywhere a team has adopted AI tools for drafting, coding, writing, or analysis.
The bottleneck Anthropic named directly is human review. Output accelerated faster than the organization’s capacity to check it. That is not a technology problem. It is a documentation and governance problem.
Anthropic’s own four category framework, trivial, routine, substantial, and poorly defined, is a useful model for any business evaluating where to apply AI internally. Work with a clear existing standard is the easiest and safest to delegate. Work without a documented process or a defined standard for success is the highest risk, regardless of how capable the AI tool is, because there is nothing to check the output against.
This is the same operational gap LiveDoc Solutions has tracked in other research this year. The MIT NANDA Initiative previously found that 95 percent of generative AI pilots fail to produce measurable business impact, largely because the operations underneath them were not documented well enough to support reliable output. Anthropic’s report describes the same failure mode from the opposite direction: an organization that documented enough to scale AI generated output, while also acknowledging that review capacity is already the limiting factor, even inside a company built to manage this.
Bottom Line
Whether AI ever begins to improve itself is a research question. Whether your operations can absorb a five times or eight times increase in output without a documented standard to check it against is a business question. It needs an answer now, not after the volume arrives.
Expert Tip
Before assigning any workflow to an AI tool, write down the standard the output must meet and who is responsible for checking it. If that standard does not exist yet, documenting it is the actual first step, not the AI tool.
Ready to Close the Gap
LiveDoc Solutions builds the documentation infrastructure that lets a business scale AI adoption without losing control of quality.
Click to submit an interest form today.
Source: The Batch, DeepLearning.AI, “RSI Is the New AGI,” reporting on Anthropic’s June 2026 internal productivity report. MIT NANDA Initiative pilot failure statistic previously cited by LiveDoc Solutions.
