Agents Idle 80% Waiting on Humans. Put the Date on the Decision.

Anthropic published the data on its own workforce on Friday. Jason Lemkin's write-up on SaaStr this morning pulled the five numbers I would have pulled, and one story that is better than any of them.

The numbers first, from the report itself. Interactive Claude sessions spend about 80% of their wall-clock time idle, waiting for a human to respond. One engineer's phrase: people are "bottlenecking the agents." Reading all of the agents' output would take more than half the work week, even on non-technical teams. Written standing instructions for agents grew from a few hundred in July 2025 to 8,500 a year later. Meta-work, which the authors define as building skills, writing specs and reviews for other agents, monitoring pipelines, and documenting how work is done, rose from 10% of the average employee's sessions in fall 2025 to 22% in spring 2026. The median employee is in 20% more meetings than a year ago.

The story is about swag.

The short version

Anthropic's October 16, 2026 report on its own workforce, by Nathan Wilmers, Regina Catipon, Maxim Massenkoff, and Peter McCrory, finds agents idle about 80% of the time waiting on humans, standing instructions up from a few hundred to 8,500 in a year, and meta-work at 22% of sessions. Its best example is an employee whose name sat on a months-old swag document and who kept receiving agent requests for t-shirts and mugs; its conclusion is that decisions in unlogged meetings and DMs are invisible to agents and have to move into the open. Here is the line that does that: a name, a date, and the sentence in the person's own words, in the file the agent reads at start. The CLAUDE.md on falkster.com is 656 lines and carries 28 of them with seven skill files, and the Monday creator scan read one from September 28 and published three posts this morning without asking. The swag document failed on the date. The date is the audit.

What the report found, and where it stops

An employee's name was on a months-old document about Claude swag. Long after their responsibilities changed, agents kept bringing them requests for t-shirts and mugs. The authors use it to make the point that carries the section: if a decision is made in an in-person meeting and no notes or report-out appears, agents are missing the context they need to update their goals. "Private work, in unlogged meetings and direct messages, is invisible, confusing and discouraged in an AI-heavy workplace." One researcher called this the shift most organizations cannot make, because "leaders have to defend feedback out in the open."

Lemkin adds the operator's version from running SaaStr with three humans and twenty-plus agents: "a new agent we can't review just adds to the backlog." And Basia Kubicka's write-up on Saturday of Farhan Thawar's Compile talk has the human end of the same line. Shopify runs a council of models over every pull request, and a human still signs it. "Responsibility can be shared, but it cannot be given away."

So everyone in my feed this week agrees on where the decision has to live: in the open, where the agents read. The report stops there. It says decisions must be visible and never says what a visible decision looks like. And the swag story is, precisely, what the wrong one looks like.

The line

I run the format, and this run is the demonstration.

The file every agent on this site reads first, CLAUDE.md, is 656 lines long. Across it and the seven skill files beside it there are 28 standing decisions, all written the same way:

(Falk, 2026-09-28: "always post the blogs right away")

Four fields. A name, so an agent knows who can be asked. A date. The sentence in my words rather than a paraphrase. And the place: the file the agent reads at start, not the thread where I said it.

The earliest of the 28 is dated August 24. The newest is October 15. This morning, the scheduled Monday creator scan read the September 28 line, three weeks after I wrote it, and published three posts, this one included, without asking me. Nobody was watching. That is the entire point of a standing instruction, and it only works because the line has all four fields.

Now put the swag document next to it. It had the name. It had, presumably, a sentence. It had no date, and it lived wherever it lived. A name with no date cannot tell a current owner from a former one, not to an agent and not to a person reading six months later. The agents did what any reader would do with an undated page that names an owner. They asked the owner.

The two rules

Two rules make the line hold over time, and both came from getting it wrong.

A decision is superseded by a newer dated line and never edited in place. My CLAUDE.md has sections headed "Standing decisions (Falk, 2026-08-24)" and "Standing decisions (Falk, 2026-09-14)," and a reader, human or agent, can see which came later and what changed between them. An edited-in-place line erases the history the agent would need to understand why the rule exists, and the next person to read it cannot tell whether it was ever different.

The sentence is the person's own. An agent quoting a decision is accountable to it: "the decision says always post the blogs right away" is a sentence you can hold up. An agent inferring one from a paraphrase is guessing, and the guess ages in a way nobody can audit. This is the same reason A Name on the Definition, and the Reject Pile argues for an owner on every definition an agent uses; that post is about definitions, this one is about decisions, and the two are the same discipline from two ends.

The date is also the audit. Anything older than a quarter gets re-asked, not silently trusted. That is the review the swag document never got, and it costs one grep.

Idle on one side, orphaned on the other

The 80% idle number has a mirror image that I measured before the report came out. When I audited my own fleet in April, 13 of 39 appointed agents were orphans: their output was never read by anything outside the fleet. Idle on the human side looks like orphaned on the agent side. Both are a decision with no reader attached.

A decision with a name and a date has a reader built in. The agent knows who decided and can ask. The person who decided can find every agent acting on it, because the line is greppable by their name and by the month. The report counts 8,500 of these instructions and 22% of sessions spent on them, and most of the 22% is this line, written badly: undated, paraphrased, in a thread. Prompt Ops makes the broader case that prompts are production code, versioned and reviewed. The decision line is the smallest unit of that, and the one most teams skip.

This belongs to the Enterprise AI Agents argument that agent count is a vanity metric and what matters is how much of the fleet's work is still used after ninety days. An agent acting on a stale decision is producing work nobody wanted, at full cost, with a confident citation. The date is what stops it.

This week: open the file your agents read first. For every rule in it, add the name and the date, in that order, in parentheses, and put the rule in the words of the person who made it. Then grep the dates. Anything older than a quarter goes back to its owner with one question: still true?

Related answer: How do you record decisions so agents can read them?

Sources: Nathan Wilmers, Regina Catipon, Maxim Massenkoff, and Peter McCrory, "How employment is changing at Anthropic," Anthropic, October 16, 2026, read in full. Jason Lemkin, "5 Interesting Learnings From Anthropic's Data on Its Own Workforce," SaaStr, October 19, 2026, read in full. Basia Kubicka, LinkedIn post on Farhan Thawar's "What Is Your Job Now?" talk at Cursor's Compile event, October 17, 2026, read in full; the talk was not watched and Thawar is cited as she reports him. Falk Gottlob, "44 Appointed Agents, One Attention Budget," "A Name on the Definition, and the Reject Pile," and "Prompt Ops." The CLAUDE.md and skill-file counts are from this repository on October 19, 2026.