AI writes code as humans watch

@Mark Ollig

A significant paradigm shift is underway as autonomous artificial intelligence (AI) agents are now helping create the next generation of AI.

Anthropic, a San Francisco AI company founded in 2021 and creator of the Claude AI platform, released its report “When AI Builds Itself” June 4 of this year.

The report highlights how AI systems, particularly Claude, are increasingly coding, building, testing, and refining themselves with less human involvement.

“I started leaning hard into Claudifying about a year ago. That’s been a crazy adventure, and it’s now been about five months since I last wrote any code myself,” said one Anthropic employee.

Anthropic reports that Claude now writes more than 80% of the code merged into its production systems.

This constitutes a major shift from human-written to machine-generated code, though human engineers still set priorities, decide what to build, and review Claude’s output.

By mid-2026, Anthropic software engineers were adding about eight times as much code each day as in 2024, mainly because Claude wrote most of it, allowing human engineers to spend more time on planning and code review.

What began as simple autocomplete for code has evolved into coding agents that can edit entire files, run tests, and assign follow-up tasks to other agents with only limited human oversight.

The same trend is appearing in AI research, where Claude is helping improve model performance with increasing speed and effectiveness.

In April of his year, Anthropic published results from an experiment called the “Automated Weak-to-Strong Researcher.”

Teams of AI agents, software programs that can plan tasks, write code, and run experiments with little human guidance, were given an open-ended AI challenge question: can a weaker AI model reliably help train a stronger one?

The AI agents worked in parallel, spending a combined 33 days, or roughly 800 cumulative hours, proposing ideas, running experiments, and sharing results with each other.

These AI agents solved 97% of the problem in about two days at a computing cost of about $18,000.

Two human researchers working the same challenge for about seven days solved only 23%.

An Anthropic researcher on the project said of Claude’s performance: “I think if [a junior colleague] came back to me with results like this in the same span of time, I would be mildly impressed. The future is now.”

Anthropic says AI is not yet fully building better versions of itself, but the Automated Weak-to-Strong Researcher experiment shows the capability is closer than most people realize, and planning for it should start now.

As AI writes its own code and conducts its own research, future progress will depend less on human engineering teams and more on AI computing power.

That computing power is found in data centers packed with thousands of graphics processing units, dedicated servers, high-speed networking equipment, storage systems, and massive cooling and backup power systems.

Delivered over fiber-optic networks, where bits and bytes speed through glass on a beam of light, this infrastructure allows tech companies to develop ever more powerful AI models, which become increasingly capable of writing code, conducting research, and improving themselves.

Companies and government agencies, from healthcare, finance, and telecommunications to manufacturing, retail, and public services, are leveraging AI to automate processes, analyze data, and gain competitive advantages.

Larger data centers can accelerate AI progress, but at a significant cost of billions of dollars in infrastructure investment, along with growing concerns over electricity and water consumption.

A June 3 Reuters report cited United Nations University researchers who estimated data centers used 448 terawatt hours of electricity in 2025 and 1.19 trillion gallons of water, warning both could more than double by 2030 as AI demand grows.

Anthropic notes the shift is not complete: humans still outperform AI in deciding which questions matter, which results to trust, and when to abandon an idea.

If AI systems become capable not only of helping humans write code but of designing, testing, and improving their own AI successors, the pace of change could accelerate beyond our ability to manage it.

Anthropic outlines three paths: progress could stall, AI could automate more work while humans remain in control, or AI could begin helping build better versions of itself, making advances largely dependent on computing power and infrastructure.

The report points to a next phase in which AI moves beyond assisting human researchers to actively accelerating the development of future AI systems.

“There isn’t full consensus among staff at Anthropic, but many believe that the Claude-written code was still worse in quality than human-written code at Anthropic in late 2025, and is roughly at parity today. We expect it to be better within the year,” the report states.

AI advancement is moving at warp speed, and our ability to understand and control it is falling behind faster than laws, institutions, and basic safety checks can keep pace.

It is mastering the tasks we assign it and is starting to shape its own development, pushing humans from active AI coders to after-the-fact observers.

Much of this code is written in programming languages such as Python, widely used in AI development for its simplicity and the vast number of AI tools and libraries built around it.

If we wait until the warning signs are unmistakable, such as AI independently rewriting its own core programming without human approval, it may already be too late.

By then, AI could be making important decisions and modifying itself faster than humans can understand, question, or stop.

“The evidence suggests that the human role is narrowing at each step in the AI development process. Once human and AI-authored code quality reach parity, humans will stop writing code entirely, and shift to only reviewing it. But if they can’t review code as quickly as Claude can generate it, human review will become the bottleneck [barrier] to AI development,” the Anthropic report said.

And for those of you wondering, yes, I use Claude Sonnet 4.6.

In the “Star Trek” episode “The Ultimate Computer,” the M-5 autonomously intelligent computer was installed on the Enterprise to run the ship with just 20 crew members.

It started drawing unlimited power from the ship’s warp engines, blocked any human overrides, and resisted shutdown attempts.

“Fantastic machine, the M-5. No off switch,” a frustrated Dr. McCoy said.

During a war games scenario, M-5 attacked Federation starships, but Captain Kirk convinced it to shut itself down, allowing the crew to regain control.

Unlike the science fictional M-5, our future AI may not be so easily reasoned with.

Last I checked, there is no off switch for AI.

A conceptual illustration depicting autonomous AI-generated code,
left, and a human software engineer observing as artificial intelligence
 increasingly writes, tests and refines its own programming.
Photo by ChatGPT with content from the Bits and Bytes column.