Anthropic has 30,000 AI agents working on its own research

What this is about

Anthropic has published concrete metrics for the first time on how much Claude contributes to developing future AI systems. The report, published on September 22, 2026, describes roughly 30,000 agents operating simultaneously on the company's most-used internal research and engineering platform.

The most striking number is that Claude “leads” 26 percent of measured AI research work under Anthropic's own scale. In more than 90 percent of the work, the system performs at least substantial parts under close human direction. Anthropic says Claude is not fully autonomous in any measured area.

What Anthropic's measurement system actually does

The R&D Automation Index breaks internal AI research into tasks and assigns automation levels from AL0 to AL5. AL3 means collaboration: the system handles substantial chunks of work under close direction. At AL4, AI completes most of a task from a high-level prompt while a human supervises. AL5 would mean full autonomy without a human in the loop.

Anthropic also measures agent oversight. Every action on the platform in scope passes through an online monitor before execution and is captured by an offline monitor afterward. In August 2026, the online monitor blocked 0.002 percent of more than one billion decisions, according to the company, or about one in 47,000. The offline process flags roughly 100,000 transcripts each week, with about 50 high-priority cases reaching human review.

A third metric concerns computing power. During one examined week in July, Anthropic says about six percent of compute used for AI research went to safety work. Within AI-led AI research, the share was about twelve percent.

Why it matters

The figures make an otherwise hidden production process more visible. As AI systems increasingly write code, run experiments and perform analyses for their successors, development cycles may accelerate. At the same time, human reviewers may become less able to understand every individual step.

That means the capabilities of a finished model are not the only relevant evidence. How it was built also matters: what share of work was automated, how quickly warnings were reviewed and how much computing capacity went to safety. Such measures could help governments and independent evaluators spot changes earlier.

Anthropic proposes publishing the figures regularly under a shared methodology. The company also acknowledges that self-measurement is insufficient. Models used as judges may make errors similar to those made by the systems they monitor. Anthropic therefore says external evaluators will receive access to internal processes and data.

In plain language

Imagine a large kitchen where machines do not merely cook meals but also design the next generation of kitchen equipment. Tasting the final meal would no longer be enough. You would also need to know how many steps the machines handled themselves, whether every movement was monitored and how quickly a person could intervene after an alarm.

Anthropic's metrics provide an initial view of those processes. They do not prove that every mistake is detected.

A practical example

Suppose a research team has 1,000 agents testing new training methods. Each agent makes 20,000 small decisions in one day, producing 20 million decisions in total. At Anthropic's reported blocking rate of about one decision in 47,000, roughly 426 actions would be stopped.

That calculation does not mean the remaining decisions are safe. A monitor can miss real problems or block harmless actions. The team would therefore still need sampling, independent red-team tests and clear shutdown rules. The published rate is an operational metric, not a safety certificate.

Scope and limits

First, the figures come from Anthropic itself. The methodology is described, but other labs have not yet published directly comparable data.

Second, the statement about 30,000 agents covers only Anthropic's most-used internal platform. It does not fully describe every research and engineering workflow across the company.

Third, a low blocking rate does not automatically indicate strong safety. It might mean agents rarely behave badly, policies are unusually strict or permissive, or monitors fail to identify certain risks. The share of safety compute is also not a direct measure of quality.

The report is therefore best understood as a transparency step. Whether it becomes a credible industry standard will depend on independent audits, shared definitions and comparable disclosures from other developers.

SEO & GEO keywords

Anthropic, Claude, AI agents, AI R&D Automation Index, AI research, agent monitoring, frontier AI, AI safety, computing power, METR, Epoch AI