An Error-Memory Institution in an Agent Organization: Behavioral Regularities from a 53-Day Full Audit

As large-language-model agents enter production settings, it remains empirically unclear where the reliability of agent organizations comes from. This paper reports a 53-day, write-level full audit of a production organization in which one human principal directs 26 persistent role-seats staffed entirely by AI agents; the record covers 5,130 commits, roughly 7,000 gate interceptions, and about 100 error-archive entries. Across the window, violation attempts never declined, and no model was ever fine-tuned or updated; what became reliable was not the models or the workforce but what entered history. Reliability is manufactured by two organizational-level institutions: a pre-commit gate stops violations before they are written into history, and an error archive converts treated error classes into one-shot events; rules were codified within a day, at 1.8% of write traffic in maintenance cost. The audit also distills five candidate behavioral regularities: incentive myopia, rhythm of closure, convergence by outsourcing, script-model division of labor, and sycophancy amplification. Each is paired with an adoptable institutional countermeasure. The paper has three contributions: a machine-recomputable internal audit record, five behavioral regularities with their countermeasures, and an error-memory institution whose cost and timing are measurable. To our knowledge, no prior study has observed a real multi-agent organization from the inside with a write-level audit.