At Anthropic, agents work six times their handlers' hours and wait for humans about 80% of each session

The most AI-heavy company around says its limit is no longer the agents. It is the people reading what they produce.

What happened

Anthropic published a report on how work and hiring have changed inside the company. Some of the numbers:

• By August, the agents employees deployed worked six times as many hours as their human handlers.
• Interactive Claude sessions spend around 80% of their time idle, waiting for human input.
• For non-technical staff, reading all the output of their agents would take half their work week.
• Written standing instructions for agents grew from a few hundred in July 2025 to 8,500 a year later.
• "Meta-work", such as writing specs, having one agent review another and monitoring pipelines, rose from 10% to 22% of the average employee's Claude sessions.
• The junior share of technical hires fell from 27% to 9%.

Staff built their own fixes. One engineer set up an agent chief of staff to triage agent requests. A finance manager built a purchase request review agent. TechCrunch adds that meeting time per employee is up 20% on a year ago. Anthropic itself notes it is an unusual, extremely AI-heavy company.

My take

Small businesses will hit the same wall sooner than they think. Once one automation works, the owner adds five more, and soon every one of them sends a summary, a draft or an alert to the same tired person.

How I design around that:

1. Every output ends in a decision, not a document. "Approve these 3 refunds" beats a 900 word report.
2. Batch approvals into one or two daily moments instead of a stream of pings.
3. Rank what reaches a human. Routine items get logged, exceptions get surfaced.
4. Write standing instructions down as versioned SOPs the automations read, so rules live in one place.
5. Track how long work sits waiting for a person. That number tells you where to automate next.

The win is not more agent output. It is less human reading per decision.

Approval queues, triage layers and exception reports like these are a big part of what I build for teams. You can see that work at romielwillautomate.dev.