Anthropic's October 16, 2026 report says its employees' agents now log six times their hours, yet interactive Claude sessions sit about 80% idle waiting for a person, and reading all agent output would take half a work week. Robert Iger's rule from The Ride of a Lifetime, three priorities and prompt decisions, gives the fix: sort agent requests into three lanes.
The most useful line in Anthropic's new report on its own company is not about productivity. It is an engineer comparing a fleet of Claude agents to babysitting: every child brings you a juice box with the same urgency as a cut finger. Most founders running more than two agents know the feeling; this report measures it.
The report, "How employment is changing at Anthropic," was published on October 16, 2026 by Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory. It draws on internal telemetry, calendar data, employee surveys, hiring records and 17 employee interviews. Agent hours across the company passed human hours in March 2026, and by August the agents that individual employees directed were logging six times as many hours as the employees themselves. Code commits per engineer ran at five times their pace of six months earlier, and lawyers, accounts payable and sales also processed more.
The cost shows up on the human side. Anthropic's interactive Claude sessions spend around 80% of their wall-clock time idle, waiting for a person to respond. The authors estimate that reading everything employees' agents produce would take even non-technical staff more than half their work week, at 250 words a minute. "Meta-work" (writing specs for agents, having one agent review another, monitoring pipelines) rose from 10% of interactive sessions in fall 2025 to 22% in spring 2026, and standing written instructions for agents grew from a few hundred in July 2025 to 8,500 a year later. Meetings did not shrink either: the median employee spends 20% more time in them than a year ago, and technical individual contributors 16% more.
Read it with the obvious caveat. This is a company grading its adoption of its own product, in a year when it tripled in size and sold a strong model, which makes deals easier to close. TechCrunch's October 17 write-up notes Anthropic's own admission that more code does not necessarily mean better code or business impact. The direction still matters, because the mechanics are the same at three agents as at three hundred: output scales with tokens, and review does not.
Robert Iger tells in The Ride of a Lifetime how the consultant Scott Miller cut him off as he started listing priorities: too many priorities means no priorities, so he could have only three. Iger also argues that chronic indecision does more damage than a quick decision that is occasionally wrong. An agent queue breaks both rules by default. Every request arrives looking equally urgent, and each waiting agent turns your hesitation into idle time. Anthropic's 80% figure is, in effect, a count of decisions not yet made.
The three-lane agent queue
This is the protocol I would take from the report and Iger's rule. It is not something Anthropic prescribes, and it is built for one person directing a handful of agents.
- Put your three priorities in every agent's standing instructions. Write the week's three priorities at the top, and tell each agent to tag any request it sends you with the priority it serves, or with "none."
- Lane one: blocking requests. Anything that blocks one of the three priorities gets answered within the hour, in a single sitting, not piecemeal between other work.
- Lane two: useful but not blocking. Requests that serve a priority without blocking it wait for two fixed review windows a day, and the agents batch them into one summary instead of a stream of pings.
- Lane three: "none." These get a default answer written in advance, such as "pick the reversible option and log it," and you read only the log, once a week.
- Budget the reading. If a day of agent output would take you more than an hour to read at normal speed, cut scope before you add another agent.
Where this breaks. Lane three is delegated judgment, so it belongs only on reversible choices; anything that spends money, touches a customer or uses a credential goes to lane one whatever its tag. Batching also makes the agents look worse on an idle-time dashboard. That is the trade on purpose: by Anthropic's own numbers, idle agents are cheap and unread output is not. The report adds a quieter warning. Decisions made in unlogged meetings leave agents reasoning from stale context, so if you settle something on a call, write one line where your agents can read it.
The hiring data points the same way from the outside. Anthropic's hiring managers now comment more on candidates' autonomy and judgment and less on domain expertise, and the share of junior technical hires fell from 27% to 9%. For a founder, the scarce skill this report describes is not prompting. It is deciding, quickly and in writing, what deserves your attention.
Sources: How employment is changing at Anthropic, Anthropic, October 16, 2026; Anthropic says its workforce has become less junior and is spending more time in meetings, TechCrunch, October 17, 2026; Anthropic says its AI agents now log six times more work hours than its employees, Pondero, October 18, 2026; Anthropic's needy AI agents reveal the future of work's next productivity bottleneck, Allwork.Space, October 21, 2026; How Bob Iger fought to become the CEO of Disney, Shortform (summary of The Ride of a Lifetime), checked October 27, 2026.
Related: How to Find Your Passion · Best Self-Improvement Books · How to Make Better Decisions · AI Coach App — Building It in 8 Hours
