Anthropic Is Using AI to Build AI: As Claude Leads a Quarter of R&D, RSI Is Moving from Concept to Engineering Reality

If I had to pick one recent AI development that deserves long-term attention, Anthropic’s latest disclosure about its internal R&D automation would be near the top of the list.

Not because Claude gained another feature.

Not because another benchmark record was broken.

But because Anthropic has, for the first time, put numbers around a question that has been difficult to observe from the outside:

How much of the work required to build better AI is already being done by AI itself?

The answer is striking.

By August 2026, Claude was leading 26% of Anthropic’s internal AI research and development work.

In February, that figure was below 1%.

In roughly half a year, it went from less than 1% to 26%.

The bigger story here is no longer whether AI can write code.

It is that AI is beginning to participate in a much broader loop involved in building the next generation of AI.

---

What Does 26% Actually Mean?

Anthropic uses a six-level framework, proposed by Epoch AI, to measure how automated AI research and development has become.

The levels range from AL0 to AL5.

At AL0, there is essentially no AI involvement.

At AL2, AI acts mainly as an assistant while humans still perform most of the work.

At AL3, the process becomes genuine human-AI collaboration. AI can take on substantial parts of a task, but humans still provide frequent guidance.

At AL4, however, the relationship changes.

Humans can give a high-level objective such as:

“Figure out why this training job failed and fix it.”

From there, the AI may inspect logs, diagnose the problem, modify code, run tests, examine the results, deal with follow-up issues and complete most of the workflow end to end.

The human role shifts toward supervision, correction and final approval.

This is what Anthropic describes as AI Leads.

According to Anthropic, Claude has reached this level in 26% of its measured internal AI R&D work.

More than 90% of its AI R&D has reached at least AL3, meaning AI is already acting as a meaningful research and engineering collaborator across most of the organization.

But there is another number that matters just as much:

AL5: 0%.

Anthropic has not measured any AI R&D task that is fully autonomous.

Humans have not left the loop.

That distinction is important.

---

More Than 80% of Merged Code Is Already Written by Claude

Another way to understand the shift is through code.

By May 2026, Claude was writing more than 80% of the code merged into Anthropic’s production codebase.

Before Claude Code entered research preview in early 2025, that share was still in the single digits.

By the second quarter of 2026, a typical Anthropic engineer was merging roughly eight times as much code per day as in 2024.

Anthropic itself is careful about how this number should be interpreted.

Eight times more merged code does not mean eight times more real productivity.

More code does not automatically mean better software or better research.

Still, the direction is difficult to ignore.

The cost of execution in software engineering is falling rapidly.

Work that once required an engineer to spend hours or days personally implementing a solution can increasingly look like this:

Define the goal → Claude executes → Human reviews

The engineer’s role gradually moves away from manually writing every line of code and toward defining problems, designing systems and validating results.

---

The More Important Shift: Claude Is Doing More Than Coding

If Claude were only generating code, I would not consider this development nearly as significant.

What matters more is that AI is moving deeper into the research loop itself.

Anthropic has been experimenting with Claude-based systems that can propose experiments, run them, modify code, analyze data, compare outcomes, investigate anomalies and use the results to determine what to try next.

In one AI safety research experiment, Claude-powered agents were asked to investigate whether weaker models could effectively supervise stronger ones.

Humans selected the research question and defined how success would be evaluated.

Multiple agents then generated hypotheses, ran experiments, exchanged findings and iterated on their work.

Together, the agents completed much of the experimental design and execution over roughly 800 cumulative agent hours.

Anthropic has also experimented with automated research agents for alignment work, including training models while attempting to reduce behaviors such as deception, sycophancy and privacy violations.

This represents a significant boundary shift.

In the past, AI primarily automated tasks.

Now it is beginning to automate parts of the research process itself.

Those are not the same thing.

---

Why RSI Is Suddenly Worth Taking Seriously

This brings us to a term that is appearing more often in AI discussions:

RSI — Recursive Self-Improvement.

In theory, the idea is relatively simple.

AI helps humans build better AI.

That stronger AI then improves the productivity of AI researchers.

Higher research productivity makes it possible to develop the next generation of models faster.

Those models can then further accelerate the research process.

The loop becomes:

AI → Better AI → Faster R&D → Better AI

For years, discussions of recursive self-improvement often sounded like science fiction.

The common image was of an AI somehow rewriting itself overnight and suddenly becoming vastly more capable.

Reality may look much less dramatic.

It may begin through very ordinary engineering improvements.

Today, AI writes 20% of an engineer’s code.

Then 50%.

Then 80%.

Later, it begins running experiments independently.

Then analyzing those experiments.

Eventually, it starts proposing parts of the experimental agenda itself.

There may never be one obvious “takeoff moment.”

Every individual step might look like another incremental productivity improvement.

But once those steps are connected, the underlying R&D system has changed.

Anthropic itself remains cautious about this.

The company has explicitly said that full recursive self-improvement has not been achieved and that such an outcome is not inevitable.

I think that distinction is essential.

What we are seeing today is not full RSI.

A more accurate description is:

Some of the building blocks required for RSI are beginning to appear inside real production and research environments.

---

The Scarce Resource May Shift from Execution to Judgment

There is another consequence that is easy to overlook.

As coding, experimentation and data analysis become easier to automate, the scarce resource in R&D may move upward.

A strong researcher today still spends a large amount of time writing code, debugging systems, processing data, running experiments and organizing results.

If agents increasingly handle that execution layer, the difficult questions become different:

What should we research?

Which problem is actually worth solving?

Can we trust the experimental result?

Did the model discover a shortcut rather than solve the intended problem?

What should we test next?

When should we stop?

In other words, the better AI becomes at doing things, the more human value may concentrate around:

defining goals, asking the right questions, setting constraints and making final judgments.

From this perspective, AI may change not only the size of research teams, but also their organizational structure.

A future researcher may not manage several junior engineers.

They may supervise a fleet of always-on research agents.

---

“Will AI Replace Programmers?” May Be the Wrong Question

For the past two years, one question has appeared constantly:

Will AI replace programmers?

What is happening inside Anthropic makes me think this framing may be too simplistic.

A more useful question might be:

What does an engineer become when they can direct 20, 100 or even 1,000 AI agents?

If one engineer could previously run only a few experiments per day, that same person may eventually supervise dozens in parallel.

If a research team once needed dozens of people to handle implementation work, a much smaller human group may eventually coordinate a large agent workforce.

At that point, the change is no longer simply that “one engineer becomes 30% more productive.”

The fundamental unit of R&D itself begins to change.

Historically, that unit was the individual human.

In the future, it may increasingly become:

Human + Agent Fleet

---

What I Find More Interesting Is Why Anthropic Is Publishing These Numbers Now

There is another important aspect to Anthropic’s announcement.

The company is not only demonstrating Claude’s capabilities.

It is also trying to establish a way to measure how quickly AI is automating AI research itself.

How much work is AI-led?

How much has reached genuine human-AI collaboration?

How much still depends primarily on humans?

Has any task reached full autonomy?

Anthropic has suggested that frontier AI labs could eventually publish similar metrics on a regular basis and potentially subject them to third-party verification.

That could make AI R&D automation observable and comparable across organizations.

I think that idea may ultimately matter more than any one percentage.

In the future, the pace of AI development may not be determined only by how much the next model improves on a benchmark.

There may be another metric that matters just as much:

How quickly is AI becoming capable of accelerating its own development?

If that 26% eventually becomes 50%, then 70%, or if genuine AL5 tasks begin to appear, those changes may be more important than another model launch.

---

AI Plus Lab Observation

I do not think we have reached the stage where AI is independently developing the next generation of AI.

Anthropic’s own data still shows humans controlling several critical parts of the process:

selecting research directions, defining goals, creating evaluation criteria, supervising execution and deciding whether the results should be accepted.

But the direction is becoming increasingly clear.

AI R&D is moving from:

Humans conduct research, while AI provides assistance

toward:

Humans define the direction, while AI performs an increasing share of the research process

If this continues, the next phase of AI competition may no longer be defined only by whether GPT, Claude or Gemini leads a particular benchmark.

A more important question may be:

Which AI lab will first build an AI research system capable of continuously accelerating the development of its own next-generation models?

That may be where the nature of the AI race really begins to change.

For years, we have been asking:

How powerful can humans make AI?

The next question may be:

How fast can AI help humans build better AI?