Your coding agent just shipped another commit. Did anyone actually review it? Anthropic admitted this week that over 80% of the code merged into its own codebase is now written by Claude. Engineers are shipping eight times as much code per quarter as they did before 2025. That's not a productivity win. That's a control problem.

The numbers that matter

Anthropic's internal metrics paint a picture of AI eating its own tail. The company's "Claude Code" system now handles the entire development loop: writing, testing, evaluating, and sometimes debugging its own code. The time AI takes to complete tasks autonomously doubles every four months. In May 2025, Claude delivered a 3x speedup over human-written baseline code. By April 2026, that jumped to 52x.

The company's own engineers report that Claude-written code was "somewhat worse" than human code in late 2025. Today it's roughly at parity. Anthropic expects it to be "strictly better" within a year.

In a recent safety research experiment, Claude-powered agents recovered 97% of the performance gap on an AI safety problem. Human researchers recovered 23% over the same period. The model's ability to choose the correct investigative path improved from 51% in November 2025 to 64% in April 2026.

Standard benchmarks like SWE-bench (software engineering) and CORE-Bench (research reproduction) have been "saturated" by AI models in roughly two years. The field is moving faster than the ability to measure it.

How we got here

Anthropic traces the evolution in four phases. From 2021 to 2023, humans did all the coding. Between 2023 and 2025, chatbots helped with snippets. Then coding agents began writing entire files independently. Now autonomous agents run code and delegate tasks to other agents.

The next step is obvious: agents training and building their own successors. That's recursive self-improvement, and Anthropic says it hasn't crossed that threshold yet. But the trajectory is clear.

One Anthropic employee described the shift: "I started leaning hard into Claudifying about a year ago. That's been a crazy adventure and it's now been five months since I last wrote any code myself." Another said the shape of work today is "humans have ideas, and the models implement, test and evaluate them an order of magnitude faster than before."

The safety case

Anthropic is calling for a global mechanism to pause frontier AI development. They want an arms-control-style regime where multiple labs, including OpenAI, Google, xAI, and Meta, agree to slow or temporarily halt development. The goal: give society time to adapt.

The company argues that unilateral pauses won't work. If one lab stops, competitors race ahead. The only credible path is coordination. They plan to spend the coming months convening policymakers, researchers, and rivals.

The proposal comes with a conflict of interest. Anthropic is days away from a potential IPO that could value the company near $1 trillion. Critics accuse the company of using safety rhetoric to slow competitors while it continues to build. David Sacks, a Trump adviser, called it regulatory capture. The company denies this.

Anthropic also recently weakened its "Responsible Scaling Policy." In February 2026, it scrapped its commitment to never train a system without guaranteed safety measures, citing competitive pressure. That makes the current call for a global pause feel inconsistent.

Community reaction

The Hacker News discussion captured the tension. One commenter noted that if AI can build itself, "it can run companies of size 3000 with a few people." Another saw a pattern: predict AGI, claim job replacement, announce dangerous models, pivot to "AI building AI" to drive IPO interest.

There's skepticism about whether a for-profit entity can prioritize safety over investor returns. The community remains divided on whether Anthropic's goals are altruistic or a calculated moat-building exercise.

On Reddit, the reaction was more blunt. "Anthropic claims AI will be able to improve itself without human intervention. Wake me up when there's third-party proof," one user wrote.

Sources

So What

The 80% code authorship number is the one that matters. Not because it means AI is replacing developers, but because it means the people building these systems are already trusting them to build the next version. The safety research experiment where Claude recovered 97% of the performance gap versus 23% for humans is the real proof point. That's not a tool anymore. That's a colleague.

Anthropic's call for a global pause is either the most responsible thing a frontier lab has ever done, or the most cynical IPO play. Probably both. The company weakened its safety policy months ago because it couldn't afford to fall behind. Now it wants everyone else to slow down. The conflict is baked in.

What matters isn't whether Anthropic is sincere. It's whether the rest of the industry will listen. The answer, based on the competitive pressure and the looming IPOs, is probably not. And that's the real problem.