Anthropic Says Claude Leads 26% of Its AI R&D. “Leads” Needs Some Explaining.

That is less dramatic than a self-improving machine running the lab. It is also much closer to something an engineering team could use, which makes it more interesting to me.

The percentage comes from Anthropic’s R&D Automation Index. The company is trying to measure how much of its work has moved from no AI involvement through assistance, collaboration, leadership, and finally full autonomy. Claude reaches the “leads” level for 26% of the measured work and at least “collaborates” on more than 90%. Anthropic says none of the measured work is fully autonomous.

Those words need more attention than the percentages.

Supervision is not a permissions specification

Anthropic uses a six-level scale adapted from Epoch AI. At the level it calls “collaborates,” an engineer stays involved while Claude works through a task. When something unexpected turns up, Claude may stop and ask the engineer what to do.

At “leads,” the engineer can hand Claude a high-level request and leave it to investigate, implement, test, and handle the surprises. Anthropic’s example is a failed nightly data pipeline. Claude receives the alert, finds the bad stage, writes the fix, tests it against a copy of the data, compares the result with the last good run, and documents what changed.

In that example, the agent stops short of deployment and a human decides whether it ships.

The example describes a human change gate. It does not establish that every agent lacks production credentials or that every task has the same deployment controls. I would check those separately: a workflow that asks an agent to wait for approval is different from a service account that cannot deploy without it.

The distinction matters beyond terminology. A team using an agent this way needs good alerts, usable logs, representative test data, a way to compare outputs, source control, and a deployment process that can accept or reject the proposed change. If those systems are weak, the agent does not make them disappear. It reaches them faster.

Claude helped calculate how much work Claude performs

Anthropic did more than ask employees whether AI makes them productive. It sampled 20% of the staff in each department involved in model R&D for every week in July 2026. A Claude research agent reviewed internal work records and produced a list of roughly 15,000 tasks. Claude then organized them into a tree containing 378 specific leaf categories, including incident postmortems, reinforcement-learning network policy, and evaluation-platform defect work.

Another Claude agent researched how each category was performed. An independent Claude judge assigned the automation level. Each sampled person-week had equal weight, divided equally among that person’s listed tasks. This approximates person-time; it is not a timesheet measurement. A five-minute fix and a multiday investigation in the same list receive equal shares.

This is a serious attempt to measure a messy system. It is not an independent measurement.

Anthropic is using Claude to inspect internal records, classify work, and judge how much of that work Claude performs. The company checked those ratings against people responsible for the work. It reports 59% exact agreement between the model and human ratings, compared with 35% exact agreement between human raters, and 97% agreement within one level.

That is useful evidence that the number was not simply invented by a marketing department. It does not resolve the circularity. The task inventory, supporting evidence, category boundaries, and final ratings all pass through Anthropic’s own systems. Anthropic acknowledges this and says third-party verification and a common methodology would be needed before comparing labs.

There is another limitation hidden in any automation percentage: the basket of work is frozen. The index can show that Claude is doing more of the jobs people performed in July. It can miss new work created by the automation itself. Faster implementation may produce more experiments, more infrastructure, more reviews, and entirely new failure modes. Automating 26% of yesterday’s work does not mean 26% of the people are now unnecessary.

The 30,000-agent operation is the bigger systems story

The same report says approximately 30,000 agents were doing research and engineering work concurrently on Anthropic’s most-used internal platform in August. That number surprised me more than the 26%.

Anthropic says every action on that platform passes through an online monitor before execution and is also ingested by an offline monitor afterward. The online system blocked about one in 47,000 decisions during August. The offline system flags roughly 100,000 transcripts each week for additional automated review, with approximately 50 high-priority cases reaching humans.

The agents have persistent identities that are separate from the model version running them. Their messages are shared and linked to original material rather than passed privately from one agent to another. Actions and communications can therefore be traced back to a specific agent even after the underlying model changes.

That is an operational design, not a chat interface.

Anyone planning to put a large number of agents into real engineering work eventually runs into the same nearby systems: identity, authorization, audit records, secrets, network access, artifact provenance, monitoring, escalation, retention, and incident response. An agent that can write a correct patch but cannot be tied to the inputs, tools, approvals, and messages that produced it is not ready for important work.

The blocking rate describes how often the monitor intervened, not how many harmful actions it successfully detected. Without known-bad test cases and missed-action measurements, a low rate cannot establish safety. One blocked action in 47,000 also becomes substantial work when the system performs more than a billion decisions in a month. Rare behavior becomes routine operational volume at that scale. Anthropic’s numbers imply tens of thousands of interventions, followed by a separate pipeline that reduces a much larger transcript-review queue to about 50 cases for people. The safety system is itself a large automated system that needs testing and oversight.

Engineering changes before headcount does

I do not think the immediate lesson is that 26% of AI researchers have been replaced. The measurement is based on categories of work and employee time, not completed jobs, business value, staffing, or defects prevented. “Leads” explicitly retains a human owner.

The nearer change is how engineering work is packaged and reviewed. A person can hand an agent a larger unit of work, but only if the surrounding system makes the result inspectable. The bottleneck moves from typing the change toward defining the task, supplying access, evaluating the evidence, and deciding whether to deploy.

That can be a real productivity gain. It can also flood a team with changes that are individually plausible and collectively difficult to understand. More output increases the importance of architecture, test quality, review capacity, and knowing who owns the result after it ships.

Anthropic’s 26% figure is a prototype internal metric published by a company that benefits from demonstrating rapid automation. I would not use it to predict when AI replaces engineers, and it cannot yet be compared cleanly with another lab. I would use it as evidence that frontier AI development has already moved beyond occasional chatbot assistance into managed fleets of agents doing substantial pieces of engineering work.

For a team adopting this approach, I would start with one concrete check: can the agent’s credentials deploy a change that its reviewer has rejected? Test that in an isolated environment, including alternate paths through CI and cloud APIs. A written approval requirement is only part of the answer.

Correction, October 10, 2026: The earlier version wrongly denied that Claude was helping build its successor and treated an illustrative human deployment gate as a universal production-access restriction. The report establishes neither fully autonomous successor-building nor that universal restriction. The weighting and monitoring explanations have also been clarified.