At Anthropic, AI agents wait on people 80% of the time
Anthropic's own data shows agents now produce more than people can read. Five ways to spend less attention per decision, with steps that need no new tools.
It's mid-morning. You started four agents before your first coffee. One is fixing a flaky test, one is cleaning up the Linear backlog, one is drafting release notes and one is chasing a bug a customer reported. Each of them stops to ask you something. Can it write this file? Can it merge? Does "the customer" mean the account owner or the person who filed the ticket? You answer one and two more are waiting. By lunch you've shipped a lot and read almost none of it.
Anthropic just published numbers on how its own people work with Claude. The company that builds Claude has the same problem.
What did Anthropic find?
In October, Anthropic's economics team published a study of its own workforce, built from internal agent logs, output metrics, surveys, calendars and hiring records (Anthropic).
By August, the agents employees started and steered were working six hours for every hour their people worked. Code commits per engineer were running at five times their rate of six months earlier.
The people became the slow part. About 80% of an interactive Claude session's time was spent idle, waiting for a person to answer. Reading everything their agents wrote would take even non-technical staff more than half their work week.
One engineer compared it to babysitting, where a child who wants a juice box opened sounds as alarmed as one who has cut a finger. Another described "an army of Claudes, with every one asking for my attention."
Staff responded by building tools to manage the agents. The report calls this meta-work, and it grew from 10% of the average employee's interactive Claude sessions in fall 2025 to 22% in spring 2026. Written standing instructions for agents went from a few hundred to 8,500 in a year.
Missing context still caused trouble. One person's name sat on a months-old document about Claude swag, so agents kept sending them requests for t-shirts and mugs long after their responsibilities had changed.
IBM researchers who interviewed developers call the result vibing fatigue, and they point to the stream of permission requests as one cause, on top of the cost of switching between tasks (IBM). You don't need Anthropic's budget to fix most of it. It comes down to how many decisions reach you and how much each one costs.
Where does your attention actually go?
Before changing anything, count for one normal day. Keep a tally of every time an agent stopped for you, what it asked, and whether your answer was ever anything but yes.
Most people find their asks fall into a few groups. There are routine reads and edits you always approve. There are questions the agent could have answered from a doc. And there's a small number of calls that need you, like a merge to main, an email to a customer or a payment.
That last group is the only one that should reach you while you're in the middle of something.
Step 1. Write down what never needs you
Anthropic's 8,500 standing instructions are this step at scale. Every rule you write is a question no agent asks again.
Start with the always-yes list from your tally. Most coding agents read a project instructions file, such as CLAUDE.md for Claude Code and AGENTS.md for Codex, and have their own allow lists for commands and tools. Put routine reads, test runs and edits inside the project on those lists.
Then write the short list of things that always need a person. Sending anything to someone outside the team. Merging to main. Deleting data. Spending money. Keep it short enough to remember.
Step 2. Make every question cheap to answer
A question that makes you open three tabs costs ten minutes. One that gives you the options and the agent's pick costs ten seconds.
In the same instructions file, ask your agents to bring decisions with two or three options, the one they recommend and what happens next on each. "Merge now, or wait for Dana's review? I recommend waiting, since it touches billing." You can answer that from your phone in a hallway.
Step 3. Read finished work in batches
Answering a question that blocks an agent and reading work it finished are different jobs. Mixing them is how a morning disappears.
Answer blocking questions as they come, from one place. Read finished work at set times, once or twice a day. One developer goes further and folds even blocked prompts into a single brief they read from 8:00 to 8:30 each morning (abdur.ai). And if finished changes pile up faster than you can read them, stop starting new ones until the queue drains (Promptly).
Step 4. Keep decisions where agents can read them
The swag story is funny until it's your production database. Agents act on whatever context they can find, and a decision made in a meeting or a DM doesn't exist for them.
After any decision that changes what an agent should do, write it down where agents look, in the instructions file, the ticket or a shared doc. When someone's role changes, update the docs and tickets that name them. Anthropic's researchers concluded that working and deciding in the open is critical in a company this heavy on AI, because agents can only reason from what was written down.
Step 5. See every waiting question in one place
Agents in different apps ask in different places. A Claude Code prompt sits in a terminal, a ChatGPT action waits in a browser tab, and a Cursor question waits in an editor you closed an hour ago. The question you never saw costs the most, because the agent sits idle until you find it.
Pick one place where every agent's questions land, whether that's a channel, a board or a tool built for it, and make it the only place you check between batches.
Where Kin helps today
Kin sits between your AI apps and the accounts they act on. Every call goes through one address, Kin checks your rule for that action and account, and it records what happened. It works with the AI apps you already use and any MCP server.
Rules do the work of step 1. For each action and account you choose Allow, Ask me or Block, and Kin checks your rule on its server on every call. You can also set a whole connection to ask only before sending. Routine calls run without reaching you, and a send or a delete waits for you. Read more about rules.
Every request that does need you lands in one list, which covers step 5. The agent that's asking is named first, and requests sit under the task they belong to. You can answer on the web, your phone or your Watch, and add a note that goes back to the agent as your instruction. When an app sends more than 10 requests in a minute, Kin puts the rest in one hold that you settle with one answer. See approvals and approving Claude Code from your phone.
Agents can also ask you a question with options and their recommendation, which is step 2 built in.
On Kin's paid plans, work can be shared across your AI apps as tasks with progress, handoffs and notes. Decisions live in the task, where the next agent reads them, so step 4 happens as part of the work.
Activity shows every call through Kin with the AI app, the account, the action and the outcome, and it names the coding agent and model behind the call. When you sit down to read, you start from one record.
Your next step
Run the one-day tally from the start of this post. It takes no tools, and it'll show you which step to take first. If most of what reaches you is routine, create a free Kin account, connect one AI app, set its connections to ask before sending, and see how much stops reaching you after a day of normal work.
Sources
Questions
How much of an AI agent's time is spent waiting for a person?
In Anthropic's October 2026 study of its own workforce, about 80% of an interactive Claude session's time was spent idle, waiting for human input.
How many AI agents can one person manage at once?
There's no fixed number. Watch how long finished work waits for you. If that queue keeps growing, stop starting new work until it drains, and move routine approvals into written rules so fewer decisions reach you.