The Agents Worked Overtime. The Humans Went to More Meetings.
Co-Founder & CEO of CLRT
Most agent business cases rest on one quiet assumption: that the hours an agent takes over come back to the team as free time, to be spent on strategy, on customers, or simply on going home earlier. On 16 October, the company with arguably the deepest agent adoption anywhere published its own employment data, and it tests that assumption harder than any survey could. At Anthropic, the agents individual employees direct now log six hours of work for every hour those employees put in. The employees are not reclaiming their weeks. The median one spends 20% more time in meetings than a year ago, and the agents spend around four fifths of their session time waiting for someone to answer them. If you are expecting agents to give your team its calendar back, this is the report to read first.
Start with what the report is and is not. It is a vendor describing itself, built on Anthropic's internal telemetry, calendar data, surveys and seventeen employee interviews, so its figures describe one unusual company rather than the wider economy. That is precisely its value. Anthropic has moulded its workflows around agents more thoroughly than almost any organisation could, which makes it a preview of where adoption leads rather than an average of where it stands today. On its session telemetry, the combined hours worked by autonomous and interactive agents passed human working hours for the first time in March 2026, with human hours counted as headcount times a forty-hour week. By August, narrowing to the agents that individual employees direct themselves, the agents were working six times as many hours as the people handling them. On any execution ledger, that is the outcome every agent programme promises.
Then the report turns to what those hours cost the humans, and the story inverts. In interactive sessions, Anthropic measures idle time as the wall-clock time an agent spends not working, and it comes to around 80%. Four fifths of a session is an agent waiting for a person to read, decide or reply. Even for non-technical groups, reading everything their interactive sessions produce would take over half of the working week. One engineer described facing an army of Claudes, every one asking for attention, and compared it to babysitting, where a trivial request arrives with the same urgency as a real emergency. The authors state the conclusion plainly: the acceleration makes human attention the bottleneck for agent output. Execution became abundant almost overnight. The scarce input moved to the person whose answer every agent is waiting for, and that person did not get a second pair of eyes.
Where did the hours go instead? Partly into rooms. On calendar data, the median employee spends 20% more time in meetings than a year ago. The report is candid that some of this reflects a company that tripled in size and shifted toward meeting-heavy roles, but the rise holds among technical individual contributors, the staff who work most directly with agents, whose meeting time is up 16%. The interviews explain why. Chasing down a bug increasingly uncovers organisational and systemic problems that need people to convene and decide, and strategic questions still require debate. In one large department's Slack channels, agents did not replace human conversation. They added to it, layering agent replies on top of an unchanged volume of human ones. The report's own line is the sharpest in it: a break from managing the agents typically looks like jumping into a meeting.
The rest of the week went into work about the work. Anthropic classifies a share of interactive sessions as meta-work: writing plans and specs for other agents, building dashboards to monitor pipelines, setting one agent to review another's output. That share rose from 10% of the average employee's sessions in autumn 2025 to 22% by spring 2026. Written standing instructions for agents grew from a few hundred to 8,500 in a year. In an April survey, 20% of the employees who answered described their role shifting from doing the work to managing Claudes. None of this appears in a business case, because a business case counts the task the agent performs and stops there. What the agent actually does to the calendar is convert execution time into attention time, coordination time and supervision time, three categories nobody budgets because nobody measured them before.
For a leader planning an agent programme, the lesson is not that agents disappoint. Anthropic's output figures say the opposite. The lesson is that the constraint moves, and it moves onto the most expensive people in the building. Every agent switched on is another claimant on someone's attention, and the claim scales with agent output, not with headcount. There is a quieter cost as well. Agents reason from what they can see, so decisions made in unrecorded meetings leave them working from stale context, and the report describes steady pressure to move human work into channels the agents can read. That is a change in how a leadership team operates, not a software setting. The organisations that gain most from agents will be the ones that decide, before deployment, where human attention is worth spending, and design everything else to run without waiting for it.
Agents do not give a team its week back. They convert execution time into attention and coordination time, and nobody budgets for either.
A deeper dive
The mechanism is a queue, and queues behave in ways intuition does not. An agent that finishes its part of a task in minutes and then waits for a person has not saved that person any time; it has moved them onto the critical path. With one agent the wait is barely noticeable. With a dozen, each finishing, asking and waiting independently, the person becomes a single server facing many requesters, and an idle share of around 80% is what that looks like from the agent's side. Adding agents does not shorten the queue. It lengthens it, and because each request arrives with equal urgency, the human spends attention triaging rather than judging. Coordination compounds the same way. Agents working at volume surface more exceptions, more contradictions between systems and more decisions that cut across teams, and those are exactly the questions that send people into a room. A firm reading its agent dashboards will see throughput climbing. The cost is recorded somewhere else, in calendars and in the response time of senior staff, where no dashboard is looking.
The second-order trap is assuming your organisation will reach this wall later than Anthropic did because it is adopting more cautiously. It is more likely to reach it sooner. In a companion study published the same day, also vendor research, 59% of an unweighted poll of 1,334 Anthropic staff said AI cuts a suitable task to a quarter of the time or less, against 26% of US adults who use AI at work in a weighted sample of 10,980. The builder's fluency is not your team's. Anthropic also absorbed its agent volume with thousands of standing instructions, purpose-built monitoring and a culture that pushes decisions into the open. A firm without that scaffolding gets less agent output through each senior hour, so the attention bottleneck binds at a far lower number of agents. Plans priced on the vendor's week overstate the hours coming back and understate the attention going out. The honest planning unit for an agent programme is not hours saved. It is senior attention consumed per unit of agent output, and almost nobody measures it.
Work with CLRT
This is the problem CLRT works on. We do not treat the number of agents switched on as progress. We map where an agent programme will spend your senior people's attention before it spends it: which workflows can run to completion without waiting on a person, which decisions genuinely need convening, and how much review and coordination each agent will draw, and from whom. That judgement decides where agents should be pointed in the first place. If your teams have more agents and fuller calendars, start with the CLRT Ascent diagnostic at ascent.clrtstudio.com, then talk to CLRT about the attention budget your plan is missing.

Vishal Sachar is the Co-Founder and CEO of CLRT in Dubai. He helps businesses put agentic AI to work, leads practical AI workshops for leadership teams and operators, and writes on agentic systems, AI governance and the economics of automation. Reach him at vishal@clrtstudio.com or on LinkedIn.


