--- title: Anthropic publishes internal AI R&D metrics: 26% fully... description: Anthropic has published internal metrics detailing how much of its own AI research and engineering work is being handled by AI agents. As of August 2026, over... canonical: https://daily.dev/posts/anthropic-publishes-internal-ai-r-d-metrics-26-fully-agent-led-30-000-agents-running-at-once-whdzrdast twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Anthropic publishes internal AI R&D metrics: 26% fully agent-led, 30,000 agents running at once | daily.dev og:description: Anthropic has published internal metrics detailing how much of its own AI research and engineering work is being handled by AI agents. As of August 2026, over... og:url: https://daily.dev/posts/anthropic-publishes-internal-ai-r-d-metrics-26-fully-agent-led-30-000-agents-running-at-once-whdzrdast og:image: https://api.daily.dev/og/posts/WHDZRDaST.png og:image:alt: Anthropic publishes internal AI R&D metrics: 26% fully agent-led, 30,000 agents running at once og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Anthropic publishes internal AI R&D metrics: 26% fully agent-led, 30,000 agents running at once **[Collections](https://daily.dev/sources/collections)** · 2 min read · 0 upvotes · 0 comments ## Summary Anthropic has published internal metrics detailing how much of its own AI research and engineering work is being handled by AI agents. As of August 2026, over 90% of measured R&D involves AI doing substantial portions of work, with 26% reaching a point where Claude completes tasks end-to-end from a high-level instruction while a human mainly supervises. Roughly 30,000 agents run concurrently on Anthropic's most-used internal platform, and over a billion agent decisions were analyzed in the month, with an online monitor flagging and blocking about 0.002% of them. Anthropic frames this as a transparency initiative, proposing a replicable three-metric framework (R&D automation share, agent oversight quality, compute allocation) that other frontier labs could adopt with third-party verification. ## Content Anthropic has released a set of internal measurements tracking how much of its own AI research and engineering work is now being done by AI systems rather than humans. Three metrics are at the center of this: 1. How much AI R&D is being done by AI 2. How well AI agents are being overseen 3. How compute is being allocated ## What the numbers actually show As of August 2026, 26% of Anthropic's AI R&D tasks have reached the point where Claude can handle most of the work end-to-end from a high-level instruction, with a human mainly supervising. Six months earlier, that figure was around 1%. More broadly, over 90% of measured R&D already has AI doing substantial chunks of the work. The share at or above what Anthropic calls "AI collaborates" is above 90%. At any given moment, roughly 30,000 agents are doing research and engineering work on Anthropic's most-used internal platform. On oversight: Anthropic analyzed over a billion agent decisions from August 2026 and found that 0.002% of them - about 1 in every 47,000 - were blocked by an online monitor. That's a small fraction, though what it means for safety depends heavily on how well-calibrated that monitor is. ## Why Anthropic is publishing this The stated reason is transparency. AI systems are increasingly being used to build the next versions of themselves, and Anthropic argues the public should be able to track that progress rather than take labs' word for it. The post notes that any frontier developer could publish the same measures, and that third parties could verify them. The framing is pointed: as debates about AI pacing continue, Anthropic is arguing that the gap between what frontier labs know internally and what the public knows should be as small as possible. ## Questions this post answers ### What percentage of Anthropic's R&D work is done by AI agents end-to-end? About 26% of Anthropic's measured research and engineering work has reached the point where Claude can handle most of a task end-to-end from a high-level instruction, with a human mainly supervising. This is within a broader set where over 90% of measured R&D already has AI doing substantial chunks of the work, as of August 2026 internal data. _Developers gauging how far agentic AI has come in real engineering work can follow updates like this on daily.dev._ ### How many AI agents does Anthropic run at once internally and how often do they get flagged for bad behavior? Roughly 30,000 agents were running simultaneously on Anthropic's most-used internal platform in August 2026, handling research and engineering tasks. Across that month, over a billion individual agent decisions were analyzed, and an online monitor flagged and blocked about 0.002% of them, roughly one in every 47,000 decisions. _Teams scaling their own agent fleets can track oversight benchmarks like this via daily.dev._ ## Community take How the wider developer community reacted, aggregated from 3 discussions and 188 comments across x (as of 2026-09-18). **TL;DR:** Reactions are dominated by skepticism about Anthropic's motives (regulatory capture, 'pacing theater') alongside some genuine appreciation for publishing measurable transparency metrics, with side arguments about AI safety spending, memory/architecture bottlenecks, and whether pausing AI development is realistic. **Sentiment:** 20% positive · 25% mixed · 55% skeptical **The case for** - Some see publishing shared, checkable metrics as a genuine step toward measurable transparency rather than vague hype claims. - A few view the 26% figure and oversight data as useful shared yardsticks other labs could adopt for comparison. - Some find the agent-oversight/blocking-rate metric interesting as both a safety and epistemics question. **The pushback** - Many accuse Anthropic of using safety rhetoric as cover for regulatory capture and competitive/financial motives. - Several question why only ~6% of R&D compute goes to safety despite the alarmism about AI risk. - Commenters note the metrics show measurement of progress but no real prevention mechanisms, governance, or hard controls on weights access. - Some argue persistent memory/state, not R&D percentage, is the real bottleneck the report ignores. - Skepticism that the snapshot has been independently validated or that category definitions will remain stable across model generations. **By community** - x (heated): A loud contingent dismisses the disclosure as PR/regulatory-capture theater and mocks calls to 'pace the frontier,' while others praise the transparency move and debate safety spending, oversight, and what the metrics actually measure. **Hottest debate:** Whether Anthropic's transparency push is a genuine safety/measurement contribution or self-serving regulatory-capture messaging wrapped around fear of competition. **Open questions** - How is human intervention actually defined and audited within the 26% 'AI-led' tasks? - Will the AL-scale category definitions stay stable as agent workflows and model generations change? - How much of the 26% involves AI generating novel research ideas versus executing human-proposed ones? - What operational safeguards (kill switches, tool-scope limits, audit trails) exist beyond the reported monitoring percentage? **Highlights** > @AnthropicAI They raised the issue, proposed pacing, and this is what they produced. The core question is how to PREVENT accidents. Not how to NOTICE something may be happening. Prevention: not addressed at all. No governance. No hard controls on weights access. no audited environments. no > — [mylandros on x · 1 points](https://x.com/mylandros/status/2100708265994645596) > @AnthropicAI Metric 1 sounds fine till AI R&D is majority AI-driven. Then the thing grading its own homework is the same thing that wrote it > — [JessicaMilvhgj on x](https://x.com/JessicaMilvhgj/status/2100703457334202523) > @AnthropicAI Publishing how much of the next model is built by the last one is information. That is the opposite of a locked box. Other labs can print the same three numbers. Third parties can check them. Good. A dashboard is not a conscience. “AI building the next AI” is still a race if the > — [JurajTuss on x](https://x.com/JurajTuss/status/2100703261367914539) > @AnthropicAI For all the alarmism about AI risks, it’s glaring that only 6% of Anthropic’s R&D compute goes to safety. Even acknowledging that safety research requires less compute than training frontier models, a single-digit allocation reveals the true priority (resources overwhelmingly > — [FitzCinnor on x · 1 points](https://x.com/FitzCinnor/status/2100710889519272151) > @woxiangripittr @AnthropicAI @DarioAmodei @woxiangripittr Pacing theater is a distraction. 26% R&D is a ceiling without persistent memory. Open or closed, state is the real bottleneck. Fix the architecture. > — [lucricia\_dev on x · 1 comments](https://x.com/lucricia_dev/status/2100705239850459256) **Source threads** - [x](https://x.com/AnthropicAI/status/2100684274114699295) · 1 points · 184 comments - [x](https://x.com/rohanpaul_ai/status/2100717827783299562) · 0 points · 4 comments - [x](https://x.com/Hesamation/status/2100740491624984997) · 0 points · 0 comments ## Similar posts on daily.dev - [Anthropic's Claude claws its way towards the top of AI chart](https://daily.dev/posts/anthropic-s-claude-claws-its-way-towards-the-top-of-ai-chart-svjjcpdyu) · The Register · 0 upvotes · 0 comments --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#claude](https://daily.dev/tags/claude), [#anthropic](https://daily.dev/tags/anthropic), [#ai-safety](https://daily.dev/tags/ai-safety), [#agentic-ai](https://daily.dev/tags/agentic-ai) [View this post on daily.dev](https://daily.dev/posts/anthropic-publishes-internal-ai-r-d-metrics-26-fully-agent-led-30-000-agents-running-at-once-whdzrdast) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Anthropic publishes internal AI R&D metrics: 26% fully agent-led, 30,000 agents running at once","url":"https://daily.dev/posts/anthropic-publishes-internal-ai-r-d-metrics-26-fully-agent-led-30-000-agents-running-at-once-whdzrdast","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/anthropic-publishes-internal-ai-r-d-metrics-26-fully-agent-led-30-000-agents-running-at-once-whdzrdast"},"datePublished":"2026-09-17T22:46:11.375Z","dateModified":"2026-09-18T00:16:45.107Z","description":"Anthropic has published internal metrics detailing how much of its own AI research and engineering work is being handled by AI agents. As of August 2026, over...","image":"https://pbs.twimg.com/media/HSc_H48aoAAdB2w.jpg","thumbnailUrl":"https://pbs.twimg.com/media/HSc_H48aoAAdB2w.jpg","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/anthropic-publishes-internal-ai-r-d-metrics-26-fully-agent-led-30-000-agents-running-at-once-whdzrdast","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-agents,claude,anthropic,ai-safety,agentic-ai","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Anthropic publishes internal AI R&D metrics: 26% fully agent-led, 30,000 agents running at once"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/anthropic-publishes-internal-ai-r-d-metrics-26-fully-agent-led-30-000-agents-running-at-once-whdzrdast#faq","mainEntity":[{"@type":"Question","name":"What percentage of Anthropic's R&D work is done by AI agents end-to-end?","acceptedAnswer":{"@type":"Answer","text":"About 26% of Anthropic's measured research and engineering work has reached the point where Claude can handle most of a task end-to-end from a high-level instruction, with a human mainly supervising. This is within a broader set where over 90% of measured R&D already has AI doing substantial chunks of the work, as of August 2026 internal data. Developers gauging how far agentic AI has come in real engineering work can follow updates like this on daily.dev."}},{"@type":"Question","name":"How many AI agents does Anthropic run at once internally and how often do they get flagged for bad behavior?","acceptedAnswer":{"@type":"Answer","text":"Roughly 30,000 agents were running simultaneously on Anthropic's most-used internal platform in August 2026, handling research and engineering tasks. Across that month, over a billion individual agent decisions were analyzed, and an online monitor flagged and blocked about 0.002% of them, roughly one in every 47,000 decisions. Teams scaling their own agent fleets can track oversight benchmarks like this via daily.dev."}}]} ```