How do you record decisions so agents can read them?

THE SHORT ANSWER

One line, four fields, in the file the agent reads at start: the name of the person who decided, the date, and the sentence in their own words, for example (Falk, 2026-09-28: "always post the blogs right away"). The CLAUDE.md on falkster.com is 656 lines and carries 28 lines in that shape across it and seven skill files, dated August 24 to October 15, 2026; the Monday creator scan read the September 28 line three weeks later and published three posts without asking. Anthropic's October 16, 2026 report shows why: agents idle 80% waiting on humans, standing instructions up from a few hundred to 8,500, and an employee still getting swag requests because their name sat on an undated document. Supersede with a newer dated line, never edit in place, and re-ask anything older than a quarter.

The report says decisions have to be where agents can read them and stops before saying what a readable decision looks like. The swag story is what the wrong one looks like: a name, no date. The format I run, the two rules that keep it honest over time, and how the 80% idle number mirrors the 13 of 39 orphans in my own fleet are in Agents Idle 80% Waiting on Humans. Put the Date on the Decision..

The same discipline for definitions, an owner on every term an agent uses, is A Name on the Definition, and the Reject Pile, and the charter lines for appointing an agent at all are in 44 Appointed Agents, One Attention Budget.

SOURCES

THE LONG VERSION

RELATED ANSWERS

Last reviewed 2026-10-19 · 1 min read