The 80% Problem: Governing the Human in the Loop.

Anthropic published the numbers on running itself with agents. By August, agent hours inside the company outnumbered employee hours 6 to 1. The number that stopped me is smaller: roughly 80% of the time in an interactive Claude session, the agent is idle, waiting on a person.

What Anthropic measured

The report, How employment is changing at Anthropic, came out October 16 from Anthropic's Economics team. A few of the numbers:

• Agent hours. Combined agent work hours passed human work hours for the first time in March 2026. By August, interactive agent hours outnumbered employee hours 6 to 1.
• Output. Commits per engineer reached 5x their level from six months earlier, with no difference in how often agent-written and human-written code was later reverted.
• Hiring. The junior share of technical hires fell from 27% to 9%.
• Waiting. Interactive Claude sessions spent around 80% of their time idle, waiting for human input.
• Overhead. The median employee spends 20% more time in meetings than a year ago. Meta-work, meaning building skills, reviewing agent output and monitoring pipelines, rose from 10% to 22% of sessions.

The authors estimate employees would need about half their work week just to read what their agents produce. One engineer described people as “bottlenecking the agents.”

The constraint moved to the person

If the company that builds Claude spends 80% of its session time waiting on its own people, the rest of us should plan for the same.

Writing the code got cheap. Deciding whether to accept it did not. Every approval, correction and judgment call now sits with someone, and in most companies those requests are spread across inboxes, Teams chats and whoever happened to be watching the run. Nobody can tell you how long they take, or which ones were never picked up.

Run the human step like a service

We already know how to run work that depends on a person. A help desk ticket has an owner, a priority and a clock. A claim or a purchase approval moves through a queue with an escalation path. The human in the loop deserves the same treatment.

With those five in place, the 80% becomes queue time you can measure, report on and shrink.

How Agent Mission Control handles it

We built Agent Mission Control (AMC) around this step. When an agent reaches a decision it shouldn't make alone, the run pauses and the request lands in AMC's intervention queue. Assignment rules route it to a named person in Teams, Slack, email or ServiceNow, with the context to act. The SLA clock starts, AMC tracks the request's age, and it escalates when the SLA is missed. The person approves, corrects or rejects it, and the run picks up where it stopped.

Every decision is recorded against the run and the approver, inside your own Azure tenant. Rubrics run on every workflow, and a failed run is attested with the reason instead of shipped. Because the time people spend in the queue is tracked, AMC reports cost per unit of work with the people included, not only the tokens.

Start with one question

If you have agents in production, ask this week: how long did people take to clear the requests your agents sent them last month? Who owned each one, and what happened to the ones nobody picked up?

If nobody can answer, fix that before you add another agent. More agents feeding an ungoverned queue only make the queue longer.