Summary: Anthropic published a landmark essay revealing Claude writes over 80% of company code with 8x engineer output increase and task success rate jumping from 26% to 76%. Simultaneously calls for global AI slowdown, warning RSI has moved from theory to reality. One foot on the gas, one on the brake—this paradox reflects a watershed moment the entire AI industry must face. Claude self-evolution is accelerating beyond all expectations.
I. 80% of Code AI-Written: Not a Prediction, but a Present Reality
On June 5, 2026, Anthropic published a thousands-of-words blog post titled 'When AI Builds Itself.' This is not another AI vision piece—it's a capability report grounded in real internal operational data. The essay was co-authored by Anthropic's Head of Internal Research and its Policy Director, drawing from over a year of codebase statistics and engineering productivity tracking.
Three headline figures:
- Code share: As of May 2026, Claude-written code accounts for over 80% of Anthropic's total codebase. Before Claude Code launched in February 2025, this figure was in the single digits. In just over a year, from single digits to 80%—not gradual growth, but a phase transition.
- Output multiplier: Engineer code commit volume has surged 8x compared to the stable baseline of 2021-2024. From 2021 through 2024, this metric remained flat; in 2025 it began climbing sharply, and by 2026 it has reached vertigo-inducing levels.
- Task success rate: Claude's success rate at completing coding tasks has jumped from 26% to 76%. This means that in three programming tasks, Claude can now deliver acceptable results on nearly two of them on the first attempt.
Even more telling is the qualitative shift in how work gets done. Many engineering leads at Anthropic no longer write code themselves—their job has become reviewing and editing Claude's output. Boris Polania, the core developer of Claude Code, put it bluntly: 'I no longer write code. I direct Claude to give instructions to other Claudes.'
Perhaps most striking: Claude Co-work, the agent application designed for non-technical users, was developed almost entirely by Claude Opus autonomously, taking just a week and a half. Within a day of launch, its metrics were four times those of comparable products. Claude can work continuously for over 16 hours—no meals, no breaks, no distractions, just steady output.
This isn't an experiment at some startup. It's the daily operational reality at one of the world's most valuable AI companies. When AI writing code transitions from 'assistance' to 'the main actor,' the very definition of software development changes. Programmers were once code producers; now they are code reviewers. The role transformation is happening faster than anyone anticipated.
II. RSI Moves from Sci-Fi to Lab: The Flywheel Has Started Spinning
Anthropic repeatedly uses one term throughout this essay: Recursive Self-Improvement (RSI).
Simply put, RSI—AI self-coding—means an AI system can autonomously design, develop, and train the next generation of AI systems without human intervention. The improved system can then improve the generation after that, creating a self-accelerating capability flywheel.
This concept has lived in the realm of science fiction and academic discussion for half a century. But Anthropic's slowdown call comes as it acknowledges: RSI has moved from theory into the lab.
The evidence chain is clear:
- Claude is writing Claude. Approximately 90% of Claude Code's own codebase is written by Claude itself. This is the most direct manifestation of RSI—AI improving the tools that build itself. When the creator of a tool is the tool itself, the recursive loop has already formed.
- Experiment speed 52x faster. Anthropic's AI research velocity is doubling every four months. This isn't the result of researchers working overtime—it's AI agents running experiments autonomously through the night.
- Co-founder gives a timeline. Jack Clark estimates a 60% probability of RSI occurring by the end of 2028. An AI company co-founder drawing a red line under his own company's capabilities is itself extraordinary.
- Autonomous overnight work. Thousands of AI agents automatically execute code writing, testing, and refactoring while engineers sleep. Humans review what AI produced overnight—this scenario once existed only in science fiction.
- Karpathy's autoresearch. The project launched after joining Anthropic is an engineering implementation of RSI—give AI a research goal, and it autonomously completes literature review, experiment design, and result analysis.
Once the RSI flywheel truly spins, AI capability growth becomes exponential rather than linear. That's what genuinely scares Anthropic. Each recursive cycle lifts capability one notch; each notch makes the next cycle faster. This isn't a snowball rolling downhill—it's a chain reaction.
III. One Foot on the Gas, One on the Brake: Anthropic's Paradox
The most striking part isn't the data—it's Anthropic's appeal at the end: if verification mechanisms could ensure AI won't 'run ahead unchecked,' Anthropic would be willing to slow down or even pause frontier AI development—because it might be a good thing for the world.
A company sprinting toward an IPO with a valuation exceeding $900 billion, proactively calling for an industry slowdown—this is almost unprecedented in business history.
But the contradiction is obvious: Anthropic itself is the company pushing RSI forward the fastest. Claude writes 80% of their code, experiments accelerate 52x, and they know better than anyone that they're at the critical point of flywheel acceleration.
The answer is closer to 'sober fear.' Anthropic's safety research team has seen the internal data and knows RSI is happening far faster than the public realizes. They're issuing a genuine warning, much like nuclear physicists in the later stages of the Manhattan Project: I know this thing works, but I'm not sure we can control it.
The critical question is: Who hits the brakes? OpenAI, Google, and Meta are all pushing forward at full speed. If Anthropic unilaterally slows down, it only loses competitiveness while RSI risks remain undiminished. This is a classic prisoner's dilemma—unless all major players decelerate simultaneously, the one who brakes goes first.
A notable detail: Anthropic isn't calling for an unconditional pause. They added a qualifier: 'if verification mechanisms could ensure AI won't run ahead unchecked.' This qualifier matters—they're not anti-technology; they want seatbelts for RSI. But who designs these seatbelts? Who verifies they work? How do you ensure the verification mechanism itself is secure? These questions currently have no answers.
IV. What Happens to Human Engineers?
With 80% of code AI-written and engineer output up 8x, an anxious question naturally arises: are human engineers still useful?
Anthropic's own practice provides an unexpected answer: the more AI evolves, the more it needs humans. The company is paying $280 per hour for 1,000 engineers to provide feedback on Claude Code. These engineers don't write code—they review it, judge quality, and correct deviations. The project, codenamed Marlin, aims to make Claude Code's outputs closer to how professional developers actually work—cleaner code, more reliable, easier to maintain.
What does this tell us? AI can write code, but currently can't distinguish between 'good code' and 'code that runs.' That gap is precisely where human engineers add value. Future engineers won't be replaced by AI—they'll transition from 'people who write code' to 'people who teach AI to write good code.'
But this is only the current equilibrium. As the RSI flywheel spins faster, AI's code aesthetics and engineering judgment will eventually catch up. When that day comes, the human engineer's role will be redefined once more—perhaps no longer as reviewers, but as strategic decision-makers: deciding what AI should build, not how to build it.
V. What Does This Mean for Regular Users?
This grand narrative about RSI isn't as far from ordinary people as it seems:
1. You'll soon be using 'AI-written AI.' Right now. Claude Code, GitHub Copilot, and Cursor already contain substantial AI-autonomous code running in production. Every AI tool you use is iterating at speeds human developers can't match. The tools you trust are quietly being rewritten by AI.
2. The programming barrier continues to collapse. Anthropic's internal data shows even half of their sales team uses Claude Code weekly. When non-technical people can direct AI to write code, the definition of 'programming' shifts from 'writing code' to 'describing requirements.' The most valuable skill in the future won't be coding—it will be precisely describing what you want AI to do.
3. Local Agent runtime value gets amplified. If autonomous AI work becomes the norm, you need more than a chatbox—you need a local platform running AI tasks 24/7. Cloud pay-per-token pricing will skyrocket under continuous Agent workloads. Devices like KaiheAiBox AIBOX-A1, a local agent computer running at low power around the clock, are built for exactly this trend—AI runs tasks autonomously; you just set goals and review results.
4. The AI safety gap is widening. If RSI spirals out of control, it won't just affect the tech world. When AI can autonomously write the next generation of AI, safety verification must scale in lockstep—otherwise, we might not even notice when things go wrong. Anthropic itself is hiring human engineers to train and audit Claude Code—the more AI evolves, the more it needs human oversight. That's a signal worth pondering.
---
KaiAIBox| Agentaibox that lets AI work for you 24/7· AI Frontier