What is a knowledge harness for AI agents?

THE SHORT ANSWER

A knowledge harness is the set of rules around what an agent treats as true. The term comes from Henry Cipolla's team at Pedestal, as described by Basia Kubicka (2026-10-05): a side agent re-reads past threads for gaps and stale rules, open questions go to the named person who owns the answer, and every definition change is a pending diff that is approved, versioned, and kept readable. The diagnosis behind it is that agents rarely lack the fact; they hold several versions and cannot tell which is current. A small team can build one with a single file in version control. My reading agent's definition of relevant is one JSON file; this week it read 136 posts and kept 45, and its last change was a one-line diff with a date, a name, and the post that proved it. I add one rule: read what the definition rejected, because that is where every bug in mine was found.

A Name on the Definition, and the Reject Pile has the full argument. Kubicka and Cipolla's three rules all end with a person's name on a definition, and the reason is the diagnosis: a wrong answer usually means several versions of a fact, not a missing one.

What I can add from my own work is small and concrete. The definition file behind my reading agent has a commit history, and the first-day bugs written up in I Automated My Reading. The Filter Was Wrong Three Times. were all found in the discard pile. A keep list is selected to look right. The same failure from the input side is in Every Check Was Green. Four Sources Were Dead.. Both are part of the Enterprise AI Agents argument about which agents still do production work after ninety days.

SOURCES

THE LONG VERSION

RELATED ANSWERS

Last reviewed 2026-10-05 · 1 min read