AI is building the next AI.
Anthropic says Claude now leads a quarter of its measured research and development under human supervision. The consequential shift is inside the research process itself.
Editorial image for AI is building the next AI.VIBECIO Editorial Desk ·
The next generation of AI is being built with help from the current one. In a September 17 disclosure, Anthropic said Claude led 26% of its measured AI research and development as of August. Here, leading means handling most of a task from a high-level prompt while a person supervises. No measured category had reached full autonomy.
That puts the model on both sides of the production line: it is the product being improved and an instrument for improving the product. If agents make useful experiments cheaper to attempt, researchers can explore ideas they would previously have left untouched. If those experiments improve the next model, the research team acquires a better instrument. The possibility of compounding is compelling. Its rate, limits, and reliability still have to be established.
A factory with a feedback loop
Imagine a research team investigating why a model fails on a particular class of requests. One agent organizes the failures, another implements a candidate change, and a third prepares a comparison against the existing system. This is an illustrative workflow, not a description of an undisclosed lab deployment. Its value depends on whether the comparison answers the original question. Producing twenty plausible explanations is easy to mistake for resolving one.
Experiments also have physical and intellectual dependencies. A training run can occupy scarce accelerators long after the code is ready. Data may require inspection before it is useful. A surprising result may expose a defective measurement rather than a breakthrough. Multiplying the agents assigned to these problems does not automatically shorten every dependency. A lab can become extremely good at proposing work while its queue of unfinished experiments grows.
A better operating question is how many consequential uncertainties the team resolves per week, at what cost. An experiment that rejects an attractive idea can be valuable. An experiment that produces an impressive score on a compromised test can be worse than useless. The research system must preserve what was tried, what changed, what evidence came back, and why someone decided to continue. Otherwise, speed creates a growing archive of conclusions nobody can safely reuse.
Read the denominator
Anthropic’s figure comes from a prototype index built around a fixed basket of research tasks, weighted using an approximation of staff time. Claude helps classify the work and judge its automation level. That methodology makes the number informative about a defined production process. It does not make it a direct count of discoveries, a productivity multiplier, or a percentage of researchers who can be removed.
For an executive, those distinctions change the investment decision. Automating preparation may let a scientist spend more time choosing research directions without reducing the project’s budget. Removing a slow handoff may improve delivery more than generating additional code. New experiments may consume the capacity that automation frees. Each can be a sensible outcome, but a labor-replacement forecast cannot be read straight from an automation percentage.
The durable asset may be the accumulated research environment: usable experiment history, dependable evaluation data, reproducible jobs, and interfaces that let an agent act without reconstructing the organization from scratch. A better model can exploit that environment immediately. Without it, each model upgrade arrives with another round of improvisation. This is an architectural reason to invest in the laboratory around the model, even when the next model is purchased from someone else.
The engineering loop is getting shorter
A separate September 18 account from Google shows a more bounded version of this change. Its infrastructure team describes agents scanning code changes, specialized triage using structural evidence, and agents proposing fixes for human review. Google reports preventing hundreds of vulnerabilities per month. That is the company’s account, not an independently reproduced result. The practical detail is the placement of the work inside the development process, where a finding can still change the code being submitted.
An enterprise can apply the same design question to less exotic work. Where can a useful observation immediately produce a testable change? A recurring support failure might become a regression test. A rejected data transformation might improve a validation rule. The loop becomes valuable when the next attempt benefits from the previous result. Simply adding another conversational agent leaves that transfer of knowledge to chance.
The evidence has to escape the lab
On September 18, Anthropic announced an embedded-evaluation partnership with Accenture, led by Faculty. The proposed access reaches inside model development, beyond testing a finished release. Anthropic will fund the work directly, and it acknowledges that access and reporting standards remain unsettled. The arrangement is a commitment to build evaluation capacity, not proof that the resulting oversight is already independent or sufficient.
That creates a concrete test for the next phase: can evaluators reconstruct important decisions from evidence the developer would rather not emphasize? Aggregate success rates cannot answer that question. Failed experiments, changed evaluation criteria, and cases where a monitor missed a problem are part of the production record too. Faster research makes access to that record more valuable, because outside observers have less time to infer what changed from a polished release.
The enterprise implication is immediate even for companies that will never train a frontier model. Their suppliers are redesigning how intelligence is produced. Their own teams can redesign how improvements are discovered and retained. The advantage will come from turning additional attempts into dependable learning—and putting that learning to work before the next cycle begins.
Sources