Anthropic published a proposal on September 17 to track something that usually appears only in broad claims: how much of the development of new AI systems is already performed by AI itself. The work introduces three groups of metrics for R&D automation, agent oversight, and compute allocation, along with an initial internal snapshot from the company.

The proposal does not only measure what a model can do on a benchmark. It tries to measure the production process that builds the next model. That shift matters: capabilities can move quickly while control, audit, and transparency mechanisms remain stuck in a spreadsheet nobody opens.

Claude already leads part of R&D, but does not work alone

To estimate AI's role in its own development, the company created the Anthropic R&D Automation Index. The index organizes research and development tasks into a hierarchy and applies a scale ranging from no AI involvement to full autonomy.

For August 2026, Anthropic reports that Claude “led” 26% of measured R&D tasks, meaning it could complete most of the activity from high-level guidance while a human still supervised the work. More than 90% of the work was at the collaboration level or higher. No measured category reached full autonomy.

The numbers require careful reading. Classification is supported by Anthropic's own models, and the method uses staff time as a proxy for each task's weight. The company acknowledges these limits and argues for third-party verification and shared methods across labs. Without that, comparing company indexes would be like comparing test coverage while using different definitions of “test.”

Oversight needs to be measured as a system

The second group tracks monitoring coverage, review latency, and agent blocking or escalation rates. Anthropic reports about 30,000 concurrent agents on its most-used internal platform for research and engineering. Every action on that platform passes through online monitoring and is also ingested by offline systems.

Across more than one billion decisions analyzed during August, 0.002% were blocked by the online monitor, roughly one in 47,000. The offline monitor flagged one or two transcripts per thousand for additional review, with around 50 high-priority cases per week reaching human reviewers.

A small rate does not mean a small risk when volume is enormous. This may be the most useful part of the proposal for AI Engineering teams: coverage, review latency, and escalation are verifiable operating metrics. They reveal more about real governance than a generic claim that a system has a “human in the loop.”

Compute joins the transparency discussion

The third area measures how computing capacity is distributed. During one analyzed week, about 6% of compute assigned to AI R&D was classified as safety work. Within AI-driven AI R&D, the share was approximately 12%.

Anthropic describes these estimates as conservative and warns that compute is an imperfect proxy: safety research can require a great deal of human work while using relatively little processing. Still, tracking the same definition over time can reveal priorities and changes in direction.

The value is in the trend, not the snapshot

The publication includes unusual internal numbers, but its more important contribution is a proposal for an auditable language to track the frontier's pace. A single index can become marketing. A time series with a public method, independent validation, and cross-lab comparability can become governance infrastructure.

For companies deploying agents in production, the practical question is direct: beyond productivity, do we measure how much work is delegated, how many actions are observed before execution, and how long it takes to review a serious alert? If the answer is spread across three dashboards and two Slack threads, governance may still be more aspirational than operational.