# Agents re-read their own wrong output as ground truth

A wrong inference recorded on a persistent page becomes the starting fact for every later run.

By Ki Moon · 2026-07-20 · failure · current
Canonical: https://kihmoon.com/bench-notes/p/agents-reread-their-own-output/

**Symptom:** An agent-maintained page still states a claim about a person that a human corrected in the thread.

A prototype deal agent I built kept a persistent page per account, including a stakeholder map that later runs read as ground truth. When someone on the deal team posted an ambiguous three-word note about a buyer's new title, along with a link, the agent wrote that the buyer had left the company and cited the link even though it said otherwise.

Four defects lined up, and each one is common:

- The agent recorded its inference as fact.
- "Appears to have left" in the table hardened by the time it reached the headline.
- The run's own detector found the right record, but no stage could amend the page.
- The weekly re-read took only agent posts as input and left out the human's correction.

Correctors that only reply in the conversation never heal the artifact.

The fix is four rules, plus one for deduplication, enforced by a check before the page is written rather than by instructions alone, because a finding that contradicts today's summary is the opposite of a duplicate.

## Check your setup (read-only)

List the docs, plans or tables that 2 or more agent sessions wrote. Take the 5 most recently written claims in each one. Search the session transcripts after each claim for a human correction of the claim. Any corrected claim the doc still states in its old form is a gap.

**Passes if:** Every sampled claim that a person corrected in a later session reads as corrected in the persistent doc.

**Access needed:** local-config, agent-transcripts

## If it fails (changes to make later, not part of the check)

1. State that someone joined, left or changed role only from a primary source read in the same run.
2. Make confidence monotonic so a headline inherits the weakest hedge among the claims beneath it.
3. Rank a human correction above the agent's own earlier output. When a correction conflicts with the CRM, flag the conflict for a person. Do not overwrite the correction or the CRM value.
4. Leave a "stale, supersede at next write" marker for the next authorized writer where write ownership blocks an edit.
5. Exempt corrections from dedup. Label them as corrections to today's summary.

**Applies if:** Your transcripts show agent runs reading a page, plan or table that an earlier agent run wrote.

**Skip if:** Your job configs show every run rebuilding its output from raw sources, with no read of its earlier output.

Tags: agents, memory, data quality
