Every system I’ve ever run had a bottleneck. Find it, and you know what to fix. Ignore it, and you spend money speeding up things that were never the problem.
Anthropic just published a report that puts the bottleneck for AI agents in a very clear place. It isn’t the model. It’s us.
The numbers
The report is “How employment is changing at Anthropic” , published October 16, 2026. It looks at how the company’s own staff work now that they lean on Claude heavily. A few findings stood out to me.
By August, interactive AI agent hours outnumbered employees’ hours six to one.
Those agent sessions were about 80% idle, waiting on human input.
Reading everything the agents produced would take more than half the work week, even for non-technical staff. Their math assumes 0.75 words per token and 250 words a minute of reading.
In an April 2026 survey, asked how AI had changed their day-to-day work, 20% described their role shifting from doing the work to managing Claudes. Meta-work, like writing specs for multi-agent runs and monitoring pipelines, went from 10% of the average employee’s interactive sessions in fall 2025 to 22% in spring 2026. Written standing instructions for agents grew from a few hundred in July 2025 to 8,500 a year later. The median employee spends 20% more time in meetings than a year ago.
One engineer described “an army of Claudes, with every one asking for my attention.”
Read it like an SRE
Put on an operations hat and that 80% idle number jumps out. If a service spent 80% of its time blocked on one downstream dependency, nobody would say the service needs more CPU. They’d say the dependency is saturated.
Here the dependency is a person. Agents scale out. People don’t. Each new agent adds a little more work to the same queue, and the queue is one human with one pair of eyes.
That’s why I think human attention is the real spec for an agent system. Not tokens per second. Not context length. How many decisions per hour does this design ask a person to make, and how long does each one take?
I wrote about the on-call version of this in the pager goes off and nobody wakes up . Alerts that need a human have to be rare, specific and actionable, or people stop answering them. Agent requests are alerts. They follow the same rules.
What I’d measure
If you run agents for a team, I’d put these on a dashboard before anything about model quality.
Queue depth. How many agent requests are waiting on a person right now?
Wait time. How long does an agent sit blocked before someone answers? This is your 80%.
Decisions per person per day. Count them. If one person is approving hundreds of things, they aren’t reviewing. They’re clicking.
Bounce rate. How often does a person send work back because the agent asked without enough context to decide?
Output read vs output produced. If nobody reads most of what agents write, either the output isn’t needed or the review is fiction.
How to spend attention well
The fixes look a lot like what good ops teams already do with alerts.
Decide which decisions need a person. Some actions are cheap to undo. Some aren’t. Put the human on the irreversible ones and let the rest flow. I went into the identity side of that in who gave the agent a badge .
Write it down once. Those 8,500 standing instructions are interesting. Every instruction is a question the agent no longer has to ask. That’s attention saved on every future run. The report also mentions one engineer who set up an agent chief of staff to triage agent requests. That’s the right instinct.
Batch the asks. Ten interruptions an hour are worse than one review block with ten items. Agents can wait. Context switches cost more than the wait.
Make every ask answerable. A request for approval should carry the diff, the test result and the reason. If a person has to go dig, the agent pushed its work onto them.
Cap the fleet. If each person can usefully handle a certain number of active agents, stop there. Running more just grows the queue.
Here’s the kind of triage board I’d want on the wall for a team running agents:
WAITING ON A HUMAN 14
oldest 47 min
irreversible 3
missing context 2 (sent back)
REVIEWED TODAY 61
STANDING RULES ADDED 4
sticky note: juice box
The part that isn’t about tooling
The report is also honest about the human side. People spend more time in meetings. More of the job is coordinating, specifying and reviewing. That’s real work, and it’s tiring in a different way than writing code.
I don’t read that as a bad sign. It’s what happens any time a team gets a lot more capacity downstream of it. The job moves to deciding what to do and checking it got done. That’s been true for managers, for operators and for anyone who’s ever run a test floor.
The teams that do well with agents won’t be the ones running the most of them. They’ll be the ones that treat a person’s attention like the most expensive resource in the system, because it is.