Anthropic has pulled back the curtain on a question that has long fascinated and unsettled observers of artificial intelligence: to what extent are AI systems now building themselves? In newly published measurements, the company reveals that Claude, its flagship AI system, now "leads" 26% of Anthropic's own AI research and development work, while over 90% of all work involves AI collaboration at significant levels. The disclosure offers an unprecedented window into the pace of AI-driven automation inside a frontier lab.

The measurements, which snapshot activity from August 2026, show approximately 30,000 AI agents operating simultaneously on Anthropic's primary internal research platform—a scale of autonomous operation that underscores how rapidly AI development has shifted from human-directed to human-supervised work.

Quantifying the Self-Improvement Loop

Anthropic's disclosure centers on an "AI-led R&D Automation Index" that categorizes work using a scale developed by Epoch AI. The Automation Level (AL) framework ranges from AL0, where no AI is involved, to AL5, representing fully autonomous operation with no human oversight. Currently, no work at Anthropic reaches AL5, but the 26% at AL4—where Claude "leads" rather than merely assists—marks a significant threshold.

The company emphasizes that over 90% of its work now operates at AL3 ("collaborates") or higher, suggesting that purely human-led AI research has become the exception rather than the rule. This recursive dynamic, where AI systems accelerate the development of their successors, has long been identified as a potential inflection point in AI capability growth.

Oversight Systems for Autonomous Agents

With tens of thousands of agents operating concurrently, Anthropic has deployed two categories of monitoring systems. Online monitors operate pre-execution, blocking potentially problematic actions before they occur, while offline monitors flag concerning behavior after the fact for human review. The company notes that METR, an independent organization, has previously red-teamed Anthropic's offline monitoring platform.

The measurements also include a compute allocation snapshot from July 13-20, 2026, categorizing all computational resources across model training, customer serving, and safety research. This level of transparency into resource distribution is unusual for AI companies, which typically guard such operational details closely.

Implications for Healthcare and Beyond

For healthcare organizations deploying AI agents in clinical documentation, diagnostic support, or administrative workflows, Anthropic's framework offers a potential template for quantifying automation levels and establishing oversight mechanisms. As medical AI systems take on more autonomous roles, the measurement methodology could help health tech companies maintain regulatory compliance and patient safety standards.

Anthropic CEO Dario Amodei has called for coordination on pacing frontier AI development, and these measurements appear designed to support that goal. The company plans to embed independent evaluators from multiple organizations with access comparable to internal risk assessment teams, extending third-party verification beyond episodic audits.

The disclosures complement Anthropic's existing Responsible Scaling Policy (RSP) risk reports and Advanced AI Framework (AAIF) proposals, creating a multi-layered transparency approach as AI systems increasingly participate in their own evolution.

Sources

  1. Measurements for understanding the pace of AI development inside frontier labs — Anthropic