Anthropic runs about 30,000 AI agents on itself and blocks one action in 47,000
Anthropic has put a number on how much of its own company is run by software agents, and it is a large one. The Anthropic Institute says that as of August 2026 there were "approximately 30,000 agents doing research and engineering work at Anthropic at any one time in our most-used internal platform".
Anthropic scores its own work on an automation ladder running from AL0 to AL5. Image: Anthropic
The same post, published on September 17 by Marina Favaro and Phillie Wright, sets out what watches those agents. Every action they take passes an online monitor before it runs, "usually within seconds", and a second offline monitor reads all of it afterwards.
One action in 47,000 gets stopped
The block rate is the number that stands out. Anthropic analysed over a billion agent decisions across August 2026 and says 0.002% of them, "about 1 in 47,000", were blocked by the online monitor.
That can be read two ways, and Anthropic does not settle the argument. A block is the monitor's judgement that an action should not run, not a confirmed piece of harm, and the post does not say what any of the blocked actions were.
The offline monitor is the slower net, built for behaviour that develops gradually rather than in a single command. It ingests every action after the fact, flags "roughly one to two transcripts in every thousand" for further review, and sends the highest priority flags, about 50 a week, to a person.
Safety got about 6% of one week's compute
Anthropic also published the split it has been asked about for years: how much of its silicon goes to making the models safer rather than more capable. In a snapshot covering July 13 to 20, 2026, "about 6% of compute that went to AI R&D was allocated toward safety", rising to about 12% of the compute spent on AI-driven AI R&D.
Those two percentages cover one measured week, not the year, and Anthropic describes them as conservative estimates. The agent counts and the block rate come from a different window, August 2026, so they are snapshots of the same company taken a few weeks apart rather than one consistent picture.
All of it is self-reported. Anthropic measured its own systems with its own methodology, and no outside party has audited any of the figures. The company says it plans to bring in independent third-party evaluators with access comparable to its internal risk teams, which is the point at which numbers like these become checkable.
The agents are not running on their own
One line in the post does a lot of work: Claude "is not operating fully autonomously for any measured subset of AI R&D work". Anthropic grades its work on a ladder from AL0, no AI involvement, to AL5, fully autonomous, and nothing it measures sits at the top.
The 30,000 figure therefore describes agents working under supervision on a platform built to watch them, not a company running itself. It sits alongside the headline number from the same post, that Claude led 26% of Anthropic's AI research work in August, up from under 1% in February 2026.
For readers outside AI labs, the value sits in the shape of the disclosure rather than in any single percentage. A frontier lab has published how many agents it runs, how much of their output it reads, how rarely it steps in and how much compute it points at safety. Until those outside evaluators are in place, it is also marking its own homework.
