Anthropic said on Sept. 17 that Claude now “leads” 26% of the company’s measured artificial-intelligence research and development work, introducing a set of internal metrics that makes AI-agent oversight look less like a policy statement and more like an operating discipline.
The company said more than 90% of its measured AI R&D work involves Claude at a level where the model at least collaborates with employees. It also reported that roughly 30,000 agents were performing research and engineering work at any one time on its most-used internal platform as of August.
The Associated Press and The Washington Post independently reported the release. Anthropic emphasized that Claude is not operating fully autonomously in any of the work it measured.
The figures are self-reported and were generated partly with Anthropic’s own models. That limitation is central to the story: the company is publishing a proposed measurement system, not an independently verified scorecard.
Update: Accenture will embed evaluators at Anthropic
Updated Sept. 18, 2026.
Later Friday, Anthropic and Accenture announced a five-year partnership to place embedded evaluators inside the AI lab. The work will be led by Faculty, Accenture’s specialist AI business, and will include model evaluation and red-teaming, alignment assessments and tests of model safeguards. Each company expects to invest at least $1 billion over five years.
Anthropic said the evaluators will receive access comparable to employees, allowing them to watch models take shape, follow decisions about how they are built and deployed, and speak directly with staff. The arrangement is nonexclusive. Anthropic said it will fund Accenture’s work directly while also discussing separately funded evaluation pilots with METR and other nonprofit organizations.
TechCrunch independently reported the announcement. Anthropic also acknowledged that there are no settled standards for evaluator access, public reporting or funding. That leaves a central governance question: an evaluator can be operationally embedded while remaining financially dependent on the company it reviews.
The partnership materially advances Anthropic’s promise of third-party verification, but it does not resolve independence. For enterprise leaders, the immediate precedent is clear: model oversight is moving from periodic external review toward continuous access to systems, development decisions and incident reporting.
Anthropic is measuring how much work Claude controls
In its report, Anthropic divided AI involvement into six levels, ranging from no AI involvement to full autonomy. At the “collaborates” level, AI handles substantial portions of work under close human direction. At the “leads” level, the system can complete most of a task from a high-level prompt while a person supervises.
Anthropic used internal work records to identify roughly 15,000 granular R&D tasks and organize them into a hierarchy. Its agents then assessed how each category of work was performed, while another Claude model assigned the automation level. The company weighted categories by person-time so that frequently performed or resource-intensive work counted more heavily.
That methodology creates a baseline, but not a neutral audit. Anthropic acknowledged that a model used as the judge could repeat errors made by the system being evaluated. The company said third parties or models from other developers could eventually check the measurements.
Agent oversight is becoming an observable system
Anthropic also reported three operational measures for agent oversight: coverage, review latency and escalation rate. It said every action on the covered internal platform passes through an online monitor before execution and is ingested by an offline monitor afterward.
The company said its online system examined more than 1 billion agent decisions during August and blocked 0.002%, or about one in 47,000. The offline system flags roughly 100,000 transcripts a week for automated filtering and initial review, with about 50 high-priority cases reaching people.
Those numbers do not establish that the monitors catch every dangerous action. Coverage measures whether a monitor sees an action, not whether it correctly understands it. A low block rate can indicate well-behaved agents, weak detection or a combination of both. Anthropic said independent testing is needed to measure how often known misbehavior is caught.
The company separately examined how it allocated computing resources during one week in July. It reported that about 6% of compute used for AI R&D went to safety work, rising to about 12% for AI-driven AI R&D. Anthropic cautioned that compute is an imperfect proxy because safety research may require more human judgment than processing power.
Analysis: AI management needs reliability engineering
The strategic value of Anthropic’s report is not the precise percentage assigned to Claude. It is the attempt to define the operating telemetry required when agents perform work that once belonged entirely to employees.
Traditional software monitoring tracks availability, latency, errors and resource use. Agent operations add a different layer: how much authority a system has, which actions are observed before execution, how quickly exceptions reach people and whether monitors detect behavior they were designed to catch.
That resembles site reliability engineering more than a static AI policy. Controls must run continuously, generate comparable metrics and produce an escalation path when a system departs from expected behavior. A policy that says a person remains accountable is incomplete if no one can show which actions were reviewed, which were blocked and how long intervention took.
Anthropic’s use of AI to monitor AI also exposes a structural risk. Automated oversight is necessary at the scale the company described, because humans cannot review every action. But correlated failures become more likely when the worker, monitor and evaluator are built from closely related models. Independent tests and deterministic controls remain important because a second model is not automatically an independent judgment.
What enterprise leaders should measure now
Companies deploying agents do not need Anthropic’s scale to adopt the underlying framework. They can inventory tasks by authority level, measure what share of actions is monitored before execution, track how quickly blocked actions reach a human and test detection against known failure cases.
They should also separate assistance from control. An agent that drafts a report creates a different exposure from one that changes permissions, deploys code or moves customer data. Reporting a single adoption rate can hide the difference.
Finally, leaders should demand evidence that oversight quality improves as autonomy expands. Agent volume alone is not progress. The stronger operating measure is whether organizations can explain what agents do, intervene before irreversible actions and learn from the exceptions they detect.
Anthropic’s metrics are provisional and company-reported. Even so, they establish a practical challenge for every enterprise adopting agentic systems: if AI is becoming part of the workforce, its authority, monitoring and escalation paths must be managed with the same rigor as production infrastructure.