Anthropic shares pacing metrics, cites 30,000 agents

Anthropic's blog post lands just days after CEO Dario Amodei published a three-step plan for coordinating a slowdown in frontier model development, and the company frames the three metrics as the measurement layer that plan needs. Two of them target the specific dynamics Anthropic says worry it: frontier labs using AI to build more powerful models, and autonomous agents that delegate tasks to sub-agents and may start making consequential decisions themselves. The third treats compute as the most verifiable input to AI R&D and a possible lever in any future pacing effort.

The early readings are less dramatic than the framing. By Anthropic's own index, Claude is "not operating fully autonomously" in any subset of its R&D work. The company counts about 30,000 autonomous agents running in its computing environments — most of them on R&D — and its first compute snapshot, from July 13 to July 20, put roughly 6% of total capacity toward AI safety, plus 12% toward safety-focused AI-driven R&D. Anthropic says the combined metrics can give third-party evaluators a starting point for judging how fast AI is really accelerating.

What the disclosures actually change hinges less on Anthropic than on its rivals. Amodei's plan has drawn backing from Sam Altman, Elon Musk and Demis Hassabis, per the company, and Anthropic says it hopes to "model that transparency" and keep releasing its measurements. If other frontier labs publish comparable numbers, the gap between what labs and the public know may narrow; if they don't, these snapshots are likely to stand largely as a single-lab baseline, and the calls for a coordinated pace may stay aspirational.