What happened
Anthropic's AI Institute (published September 17, 2026, five days after CEO Dario Amodei's 'We Must Pace the Frontier' essay) disclosed three new measurement tools plus a real snapshot from inside Anthropic: how much of AI R&D is performed by AI itself, how well AI-agent actions are overseen, and how compute is allocated. The headline finding: 'Claude leads 26% of Anthropic's AI R&D work' as of August 2026 (built on Epoch AI's automation-level scale), 'The share of work at or above ‘AI collaborates’ is above 90%,' with Claude 'not operating fully autonomously for any measured subset.' On oversight, roughly 30,000 agents run concurrently on Anthropic's most-used internal platform and, over more than a billion agent decisions analyzed in August 2026, '0.002%' of actions required monitor blocking/escalation. It is the first public, method-attached 'R&D Automation Index' data on how close the field is to recursive self-improvement, and it commits Anthropic to publishing the metrics regularly with embedded third-party evaluators for verification.
Why it matters
This is the first verifiable, method-attached evidence an executive can cite on how fast AI is building itself and how well ~30,000 production agents are actually monitored — directly relevant to AI pace, oversight and third-party evaluation debates in the boardroom.
Action needed
Use these three metrics as a template when asking frontier vendors and internal AI teams for comparable transparency on AI-led R&D share, agent-oversight coverage/blocking rates, and compute allocation.