Build1 publisher3 min readPublished
LessWrong's count of 5,000-25,000 unattended AI agents rests on a share Claude guessed
LessWrong authors estimate that 5,000 to 25,000 AI agents worldwide are more than 24 hours past their last human input, using a 1-5% share that Claude guessed. Anthropic's own figures give a firmer oversight signal: about one human-reviewed case per 5 million agent decisions.
The Engineer · Build desk
What happened
- The global estimate starts from token volumes and counts 300,000 to 1,000,000 AI agent loops running at any given moment.
- Anthropic reports about 30,000 agents running concurrently, and the post guesses 2,000 to 4,500 of them are more than 24 hours past their last human input.
- From OpenAI's 3.1 agent-workdays per researcher-workday and $600 average daily inference spend, the post backs out 150 to 2,500 agents running unattended for more than 24 hours.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- contradiction Claude's guessed unattended share for Anthropic's fleet, 7 to 15 percent, sits entirely above the 1 to 5 percent applied to the world. Either the frontier labs run far ahead of everyone else, or one of the guesses is wrong.
- constraint Anthropic reviews about 50 cases a week against a billion decisions a month. An oversight plan that depends on a person seeing individual agent decisions would cover almost none of them.
- decision Anyone putting the headline count into planning or policy should use the wider 3,000 to 50,000 range until a lab publishes how long its agents run after the last human input.
The global figure comes from a chain of multiplications [2]. Each step can be checked against the post's own inputs.
1. Throughput. Google's 3.2 quadrillion tokens a month [3] is about 1.2 billion tokens a second [1]. OpenAI's API figure of 15 billion a minute [4] is 0.25 billion a second [2]. The post rounds the world up to roughly 3 billion a second [5]. That leaves about 1.5 billion a second as its allowance for Anthropic and everyone else [3]. 2. Concurrency. Take a third of that as agentic and divide by 2,000 to 5,000 tokens a second per live coding agent [6]. On those inputs as printed, the result is 200,000 to 500,000 loops [4]. The post publishes 300,000 to 1,000,000 [2]. Its cross-check, 5 million weekly coding-agent users at two hours a day and 1.5 parallel sessions [8], works out to about 625,000 [6]. 3. Duration. Apply a 1 to 5 percent share of agents past 24 hours, a figure Claude guessed [7]. The published 5,000 to 25,000 is exactly that share of 500,000. Across the post's own loop range, the same share gives 3,000 to 50,000 [5].
The first two steps rest mostly on reported volumes and a per-agent rate. The third sets the headline number. The post itself says the number of agents running autonomously past 24 hours is not reported [10]. Asking Claude is at least consulting an interested party. For the figure to hold up, someone has to measure time since last human input on a real fleet. The labs publish concurrency and agent-workdays [9][11]. Neither of those measures how long an agent has run without a human.
The lab-level guesses do not match the global share. Claude puts 7 to 15 percent of Anthropic's 30,000 concurrent agents past the 24-hour mark [9]. Even the bottom of that range is above the top of the global range [7]. Starting from OpenAI research-org spending, the post backs out 150 to 2,500 [11]. The two internal guesses sum to 2,150 to 7,000 [7], against a global floor of 5,000 [2].
The measured evidence for unattended work comes from Anthropic's own figures, as the post relays them. Anthropic logged over a billion agent decisions in August, and about 50 cases a week reached a human [12]. AI-led AL4 work means no human is needed while the task runs. It went from under 1 percent of R&D in February to 26 percent in August [13][14]. The report gives the example of fixing a broken nightly pipeline. The engineer hands Claude the alert and "wouldn't have to stay actively tuned in" [14]. If R&D staff number 1,000 to 2,000, a guessed figure, then 30,000 agents is 15 to 30 per person [15].
According to the post, OpenAI reported on September 8 that an unreleased model, running as roughly 10,000 concurrent agents, produced a counterexample to the Navier-Stokes conjecture. The run took about 88 hours [16]. The excerpt does not say whether anyone intervened during those 88 hours.
I think Anthropic's review ratio and its AL4 share make the oversight case without any global headcount. The headcount adds scale, and its range spans a factor of five because of one guessed percentage [7].
What to watch
- A lab publishing how long its agents have been running since the last human input, the figure the post says is not reported.
- Whether OpenAI's fuller account of the 88-hour Navier-Stokes run says if humans intervened while it ran.
- Anthropic's next report: whether AL4 work keeps climbing from 26% of R&D, and whether AL5 work starts.