Anthropic says Claude now leads 26% of its AI R&D as 30,000 agents work simultaneously inside the company
Anthropic says Claude now “leads” 26% of its AI research and development work, meaning the model can complete most of those tasks end-to-end from a high-level instruction while a human remains responsible for supervision and the final decision.
The figure comes from Anthropic’s newly disclosed R&D Automation Index, which attempts to measure how much of the work involved in developing frontier AI systems is itself being performed by AI.
The company also reports that more than 90% of measured AI R&D work has reached at least the “AI collaborates” level, where Claude can perform substantial portions of a task under close human direction.
As of August 2026, Anthropic says approximately 30,000 research and engineering agents were active at any one time on its most-used internal agent platform.
Those 30,000 agents should not be interpreted as 30,000 autonomous employees or as the total number of agents operating across every Anthropic system. The measurement applies to one internal platform, and Anthropic explicitly states that Claude has not reached full autonomy for any measured subset of its AI R&D work.
··········
ANTHROPIC NOW MEASURES AI R&D BY HOW MUCH OF THE TASK CLAUDE CAN CONTROL.
The 26% figure corresponds to a defined automation level rather than the percentage of research produced without humans.
Anthropic uses an automation scale originally developed by Epoch AI, ranging from AL0 to AL5.
At AL0, AI has no involvement. At AL3, described as “AI collaborates,” Claude can perform large parts of a task but remains under close human direction. At AL4, described as “AI leads,” the model can receive a relatively high-level objective, work through most of the task independently, handle intermediate problems and present the completed work for human review.
AL5 would represent full autonomy, where an AI system notices that work needs to be done, scopes the task, performs it, validates it and deploys the result without requiring human involvement. Anthropic reports no measured AI R&D category at AL5.
........
Automation level | Human-AI relationship | Practical control |
AL0 | No AI involvement | Human performs the task |
AL1–AL2 | Limited assistance | AI supports parts of the workflow |
AL3 | AI collaborates | AI performs substantial work under close human direction |
AL4 | AI leads | AI completes most of the task from a high-level prompt; human supervises |
AL5 | Fully autonomous | AI identifies, scopes, executes and deploys the work without required human involvement |
Anthropic, August 2026 | 26% at AL4; >90% at AL3 or above | No measured category has reached AL5 |
........
The change has also been rapid. Anthropic’s published chart places the share of model R&D at the “AI leads” level at below 1% in February 2026 and 26% in August.
Because the February figure is reported only as “under 1%,” an exact growth rate cannot be calculated. Data Studios can nevertheless derive two bounded comparisons: the increase is greater than 25 percentage points, and the AL4 share is now more than 26 times the February level if the reported upper bound of 1% is used as the denominator.
The measurement therefore describes a large shift in the structure of Anthropic’s internal research workflow over roughly six months, but it does not establish that 26% of Anthropic’s total corporate activity is autonomous or that Claude independently determines 26% of the company’s research agenda.
··········
THE 26% INDEX IS BUILT FROM THOUSANDS OF INTERNAL R&D TASKS, NOT EMPLOYEE SURVEYS ALONE.
Anthropic constructed a frozen map of model-development work and then assigned automation levels to weighted task categories.
The company sampled 20% of staff each week during July 2026 from departments involved in the model R&D loop. Claude research agents analyzed Slack records and other internal documentation associated with the sampled employees and produced a list of approximately 15,000 granular R&D tasks.
Those tasks were organized into a hierarchical taxonomy containing 542 nodes, including 378 leaf categories representing specific types of work. Examples given by Anthropic include evaluation-platform debugging, reinforcement-learning sandbox networking and serving-incident postmortems.
The resulting task tree was then frozen so later measurements could be compared against the same basket of work. For each category, one Claude agent investigated how the work was actually being performed across Anthropic and another Claude system independently assigned its automation level.
Categories were weighted using person-time as a proxy for economic importance. If an employee worked on four tasks during a sampled week, each task received one-quarter of that person’s weekly weight; if another employee worked on ten, each received one-tenth.
The methodology also introduces an important limitation: Claude is involved in evaluating Claude’s own role. Anthropic tested the automated ratings against assessments made independently by employees responsible for the relevant work areas.
The model and human evaluators selected exactly the same automation level 59% of the time, while two human raters agreed exactly 35% of the time in Anthropic’s comparison. Model and human ratings were within one level of each other 97% of the time.
Those figures suggest that the classification system is reasonably stable at a coarse level, but they also show that the boundary between “collaborates” and “leads” remains partly judgment-based. Anthropic itself identifies common methodology and independent verification as prerequisites before equivalent numbers from different frontier laboratories could be treated as directly comparable.
··········
30,000 CONCURRENT AGENTS REQUIRE A MONITORING SYSTEM THAT OPERATES AT MACHINE SCALE.
Anthropic says every action on the measured platform passes through online monitoring, while a second system reviews agent activity after execution.
The approximately 30,000-agent figure refers to research and engineering agents operating concurrently on Anthropic’s most-used internal platform as of August. Anthropic describes agents that can work for extended periods and can delegate work to other agents, creating an internal environment where coordination and oversight can no longer depend primarily on humans watching individual sessions.
Anthropic says 100% of actions on the measured platform pass through an online monitoring layer that can block or redirect an action, generally within seconds. The company analyzed more than one billion agent decisions during August 2026 and reports that approximately 0.002% were blocked, equivalent to roughly one decision in 47,000.
A separate offline monitoring system ingests 100% of agent activity after it occurs. Anthropic says this system initially flags approximately 100,000 transcripts per week, after which multiple classifiers remove false positives and around 50 high-priority cases per week reach human reviewers.
Data Studios calculates that those 50 human escalations represent roughly 0.05% of the initial 100,000 weekly offline flags. That calculation should not be interpreted as a failure rate for the agents: the initial flags are deliberately broad monitoring signals, and Anthropic does not state that every human escalation corresponds to confirmed harmful behavior.
........
Oversight metric | Online monitoring | Offline monitoring |
Coverage | 100% of actions before execution | 100% of actions ingested after execution |
Main purpose | Block or redirect rapidly harmful actions | Detect slower or reversible behavioral problems |
Scale reported | >1 billion decisions analyzed in August | ~100,000 transcripts flagged per week |
Escalation | ~0.002% blocked, about 1 in 47,000 | ~50 highest-priority cases reach humans each week |
Human review | Blocked actions reviewed within one week | Highest-priority cases manually reviewed |
Data Studios calculation | — | ~0.05% of initial weekly flags reach human escalation |
........
Anthropic has also changed the architecture of its internal agents to improve auditability. Individual agents receive persistent identities, allowing data and actions to be attributed to a specific agent even when the underlying model is upgraded.
Agents communicate through a shared messaging system rather than through private agent-to-agent channels. Messages can reference original material and are cross-linked with agent transcripts, allowing monitoring systems to follow chains of communication across multiple agents.
The architecture is intended to reduce correlated behavior and make one agent’s output something another agent can independently inspect rather than automatically inherit as internal state. It also creates substantially more telemetry for oversight than a conventional single-agent conversation.
··········
THE NUMBERS SHOW RAPID AUTOMATION OF AI DEVELOPMENT WITHOUT YET SHOWING RECURSIVE AUTONOMY.
Anthropic’s measurements describe a system in which AI performs increasingly large portions of model-development work while humans still retain control over consequential decisions.
The combination of 26% of R&D at the “AI leads” level, more than 90% at “AI collaborates” or above, and roughly 30,000 agents operating concurrently indicates that Anthropic has already moved beyond occasional use of coding assistants inside research teams.
AI agents are becoming part of the production infrastructure through which new AI systems are researched, tested and engineered. The measurements nevertheless stop well short of demonstrating a self-improving autonomous laboratory.
The 26% figure is a weighted automation classification, not a measure of the percentage of research discoveries generated independently by Claude. The 30,000-agent figure measures simultaneous agent instances on one internal platform, not 30,000 independent researchers with human-equivalent responsibility. Anthropic reports zero measured R&D categories at AL5 full autonomy.
There are also methodological reasons to avoid extrapolating the August snapshot mechanically. The R&D task basket is based largely on July work, automation levels are partly evaluated by Claude itself, person-time is only a proxy for the economic importance of different research activities, and the structure of AI R&D can change as humans move toward new categories of work.
The practical signal is narrower but still substantial: frontier-model development at Anthropic is already being conducted through large-scale populations of AI agents, and a material share of the underlying work can now proceed from high-level human instructions to near-complete technical outputs without continuous human intervention.
The next threshold is therefore whether the proportion of work reaching AL4 continues to rise while monitoring remains effective—and whether any important category crosses from supervised AI-led execution into AL5 autonomy, where human involvement would cease to be operationally required.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
[datastudios.org]
