Anthropic Found Humans Are the Bottleneck in Managing AI Agents. Now What?

What Anthropic reported

The source is an Anthropic Economics report, "How employment is changing at Anthropic," published October 16, 2026 by Wilmers, Catipon, Massenkoff, and McCrory. It draws on internal data on employment, AI use, output, surveys, Slack and calendar activity, hiring records, agent session telemetry, and 17 employee interviews. Fortune covered it on October 19 under a headline quoting an employee's description of "an army of Claudes" all asking for attention.

The headline numbers on agent management:

• About 80 percent of interactive Claude session time is idle, waiting for human input. One engineer said people are "bottlenecking the agents."
• Total agent work hours passed human work hours for the first time in March 2026. By August, individual employees' interactive agent hours were about six times their own.
• Meta-work, time spent managing agents rather than doing the task, rose from 10 percent of an average employee's interactive sessions in fall 2025 to 22 percent in spring 2026.
• Reading everything their agents produced would take non-technical groups over half a work week.
• In open-ended survey answers, 20 percent of respondents described shifting from doing work to managing Claudes.

Why the bottleneck moved

When one person drives one assistant, the human sets the pace and the model waits briefly between turns. When one person runs many agents, each agent's natural pace is faster than the person's ability to answer questions, approve actions, and read results. Output scales with the number of agents; attention does not.

That changes the economics. Adding a tenth agent is cheap in compute and expensive in interruptions. Past a point, more agents mean more idle agents, longer queues for the human, and less careful review of each result. The report also found the median employee spending 20 percent more time in meetings than a year earlier, so the attention budget is not growing either.

What Anthropic employees did about it

The report describes two responses worth copying.

A playbook for teams running many agents

Treat human attention as the scarce resource and design around it.

Classify decisions by reversibility. Let agents act on reversible, low-impact steps without asking. Require approval for irreversible or external actions. Most interruptions sit in the first bucket and should not reach a person.

Answer context questions with retrieval, not people. A large share of agent questions are lookups: where is the config, what did we decide last quarter, what is the naming convention. If that knowledge is indexed and retrievable, the agent finds it without waiting. Every lookup an agent can do itself is a human interruption avoided.

Make output skimmable. If reading everything takes half a week, nobody reads everything. Require a short summary, the decision needed, and the evidence, in that order. Review the summary and spot-check the evidence.

Batch approvals. A queue reviewed twice a day beats a stream of notifications. Context switching is the hidden cost of agent management.

Measure idle time. Track how long agents wait on humans and what they waited for. The most common wait reasons are the next rules to write or the next documents to index.

Caveats in the report

The authors flag limits. Tool adoption was not random, so some productivity gains may reflect who chose to adopt. Anthropic is unusual, with broad stock ownership, rapid growth, and a frontier AI culture, so results may not transfer. Several measures rely on automated classifiers, and survey wording changed over time. The direction of the finding is still hard to argue with: as agents get cheaper, the person reviewing them becomes the constraint.