After reading this, the reader knows Anthropic says Claude leads 26% of its AI R&D and explains how its internal agents are monitored.

Anthropic Measures AI-Led Research

The company proposes public metrics for automation, agent oversight and compute allocation, while acknowledging that its methodology relies on Claude.

Anthropic has published a snapshot of how AI participates in its own model research. The company says that, as of August 2026, Claude led 26% of measured AI research and development work, collaborated or led in more than 90%, and was not fully autonomous in any measured category.

The company also reports approximately 30,000 agents active at once on its most-used internal research and engineering platform. It says every action on that platform passes through an online monitor before execution and is ingested by an offline monitor afterward.

Why it matters

AI systems increasingly help build later AI systems. That can shorten experiments, write infrastructure and expand the number of research paths a laboratory can explore. It can also accelerate capability development faster than outside observers can track.

Anthropic proposes three measurements: AI’s share of R&D, agent-oversight coverage and response, and compute allocation. These measures examine the production process rather than only a finished model.

For automation, Anthropic uses a six-level scale ranging from no AI involvement to autonomous work. “AI leads” means the system performs most of a task while humans retain meaningful involvement. The company says no measured work reached full autonomy, an important limit on the headline figure.

The measurement is broad, but not independent. Anthropic sampled internal Slack and documentation, used Claude agents to organize roughly 15,000 tasks, and used another Claude judge to assign automation levels. The work map contains 542 nodes, including 378 detailed categories.

Anthropic compared model ratings with staff assessments, but a model evaluating work performed with related models can share their blind spots. The company identifies this limitation and proposes third-party verification or cross-lab evaluation with protections for confidential data.

Monitoring 30,000 agents

The oversight data describes one internal platform, not every agent Anthropic operates. Online monitors review actions before execution and can block or redirect dangerous behavior. Offline monitors look for patterns that develop more slowly and can be investigated after the event.

Anthropic says online monitoring covered 100% of agent actions and blocked about 0.002%, or roughly one in 47,000, during August. Offline systems flag around 100,000 transcripts each week for further classifier review, with approximately 50 highest-priority cases reaching human review.

Those rates require careful interpretation. A low block rate could mean agents rarely attempt harmful actions, monitors are effective at steering behavior earlier, or detection misses some events. A high number of initial offline flags may reflect conservative filters rather than actual misconduct. Without an independently labeled sample, neither number proves safety.

The company also discusses compute allocation because automation and hardware together determine development speed. In a sample week from 13 to 20 July, Anthropic classified 6% of AI R&D compute as safety work and 12% as AI-driven AI R&D. It calls the method conservative and notes that safety work is difficult to separate from general capability research.

Anthropic says it plans to embed independent evaluators with access comparable to internal risk teams. That commitment is more important than a one-time dashboard if outsiders can reproduce definitions, inspect incidents and publish disagreements.

The new metrics do not answer whether a laboratory is moving too quickly or whether its safeguards are sufficient. They provide quantities that governments, researchers and the public can debate. Consistent reporting across companies would make trends more visible; incompatible self-defined measures could instead create a false appearance of comparability.

Verification

  • Tier 1 — COMPANY-REPORTED: Anthropic reports Claude leading 26% of measured AI R&D, collaboration or higher above 90%, and no fully autonomous category as of August 2026. Source: https://www.anthropic.com/institute/measuring-pace-of-ai-development
  • Tier 1 — COMPANY-REPORTED: Anthropic reports about 30,000 concurrent agents, complete monitor coverage, a 0.002% online block rate and roughly 50 weekly human escalations. Same source.
  • Tier 1 — VERIFIED METHODOLOGY DISCLOSURE: Anthropic describes sampling internal records and using Claude agents and a Claude judge to build and rate the task map. Same source.
  • Tier 2 — ANALYSIS: Interpretations of block rates, comparability and oversight limits are editorial analysis.

Glossary candidates

  • AI-led R&D: Research work in which an AI system performs most of the task while humans retain meaningful involvement.
  • Offline monitor: A system that reviews agent activity after execution to find patterns across events.

Cold-reader sentence: Anthropic’s measurements suggest deep AI involvement in model development, but the self-evaluated methodology still needs independent auditing.