On October 16, Anthropic published a study of its own workforce, "How employment is changing at Anthropic." By August, it found, "interactive AI agent hours outnumbered employees' six to one." Its Claude sessions amounted to around 80% "idly waiting for human input."
The report's conclusion: "This acceleration makes human attention the bottleneck for agent output."
Fortune covered it on October 19 under a headline with a word for the new job: botsitting. One Anthropic employee described it as "an army of Claudes, with every one asking for my attention."
The attention problem is real. The detail worth stopping on comes earlier: Anthropic could count.
The company that builds the agents still had to build a way to see them
Anthropic knows how many hours its agents worked, how long they waited, and what share of its people's time went to managing them. That share, which the report calls "meta-work," rose from 10% to 22% of employees' interactive sessions. The report describes meta-work as "building visibility tools and monitoring agent pipelines."
For a leader of an ordinary company, that is the sentence to keep. The people who make the agents now spend about a fifth of their interactive sessions on work that includes building ways to see what the agents are doing.
Ask your own leadership team a simple question: how many agents does this company run? Then ask who switched each one on.
The agents are there. They came in through the software people already use. A sales rep turned on an AI follow-up tool in the CRM. The scheduling system added a chatbot. The bookkeeper's software started categorizing transactions on its own. A manager connected an assistant to a shared inbox. Each one was switched on by whoever found it useful, and none of them is written down anywhere the leadership team looks.
Botsitting is what that looks like from the inside. One person answers an agent here and another there, in private, and nobody else knows the work exists. Anthropic's report already named the cost of that kind of work: "Private work, in unlogged meetings and direct messages, is invisible, confusing and discouraged in an AI-heavy workplace."
You cannot hold an agent accountable that you cannot see
The usual advice about AI agents jumps straight to accountability: give every agent an owner, measure it, review it. That advice is right, and OTP was built on it. It is also one step too far ahead for most companies, because accountability starts with knowing what you are accountable for.
David Steel, who founded OTP, learned this inside his own company. His marketing agency, Sneeze It, puts its AI agents on the same org chart as its people, each with a seat and a number. Even so, in September he found that four agent seats had gone silent. They had not crashed or raised an alert. They had simply stopped doing their jobs, and for 146 days nothing in the company's system had been watching whether its agents were still producing. He found it because he asked each agent to defend its seat. The structure did not find it.
If a company that deliberately put its agents on the org chart could lose sight of four of them, a company that never wrote its agents down has no way of knowing what it is missing.
See them, seat them, meet them
The order matters. Each step depends on the one before it.
1. See them. List every agent the company runs, including the ones that arrived inside other software. For each one, write down what it touches (customers, money, data, other systems) and who switched it on. If the list is longer than anyone expected, that gap is the first useful finding.
2. Seat them. Every agent that touches customers or money gets a seat on the org chart, with a human owner and one number that says whether it is doing its job. A lead-response agent's number might be how many new leads got an answer within five minutes. A collections agent's might be the dollar amount past 60 days. If nobody will own an agent, that is a reason to switch it off.
3. Meet them. Agent numbers get read in the same weekly meeting as the people's numbers. When one goes off track, it becomes an issue with an owner and a date, and it stays on the list until it is fixed or someone deliberately lets it go.
This is what replaces botsitting. Instead of one person answering an agent one prompt at a time, each agent's work shows up as a number on a board, and a person looks at it once a week with everyone else.
What it looks like when the agents are visible
Here is one ordinary morning at Sneeze It. On October 20, before 7:36 a.m., one of the company's agents had updated all 17 numbers on its leadership board, people's and agents' alike. Another agent reported that zero human replies to its outreach had waited more than a day for a decision. A third checked every agent seat for silence and found none.
Nobody had to sit with a bot for any of it, because every one of those agents had been written down first.
Anthropic's study is about a company whose employees build the agents. The rest of the economy will meet agents differently, through the sales, scheduling, finance and service tools it already pays for. The bottleneck will be the same, and so will the fix: put people and agents on one org chart, give every seat a number it owns, and run a weekly meeting that picks up where the last one ended.
This is the seventh part of our series on agent accountability. Start from the beginning at Every AI agent got a job this fall. Who answers for what it does?