Observability and Auditability of Autonomous Agent Decisions in Production Infrastructure

Authors

  • Bharadwaj Kasinadhuni

Keywords:

agent observability; decision auditability; audit logging; large language model agents; reason-and-act loop; provenance; retention tiering; accountable automation; production infrastructure.

Abstract

Autonomous agents built on large language models are now placed in front of production infrastructure, where they read system state, select a tool, and take an action that changes something real. The logging such systems inherit was designed for deterministic services, so it records the action and its outcome. It does not record the decision. In a production partner-onboarding agent platform that has been running since 2024, with the agentic phase active for roughly twelve to eighteen months, the reasoning text produced by the reason-and-act loop exists only in process memory and is released once the action executes. The consequence is structural rather than accidental: after the fact, an operator can establish which tool ran and whether it succeeded, but cannot establish why that tool was chosen over the alternatives, because the only artefact that carried the reason was never written to storage. This paper specifies a twenty-two field decision record for autonomous agent actions, classifies every field by whether it EXISTS, is PARTIAL, or is PROPOSED against that deployment, and identifies the two insertions in a twelve-step pipeline that make the remaining ten steps auditable: structured capture of the reasoning trace before the action executes, and a full decision-record write tagged with a retention tier by risk class. Identity binding through a hashed token claim that is never user-supplied, personally identifiable information scrubbing at write time, and an environment enum gating permissible actions are described as existing properties that the added fields do not weaken. Instrumentation overhead is assessed with modelled estimates and stated derivation methods rather than measurements. The measured baseline for the deployment is 269 tool invocations over six months from previously published work; every higher volume in this paper is an explicitly labelled projection. The schema is specified and not yet implemented, so no evaluation of it is claimed.

Downloads

Published

2026-09-28

How to Cite

Kasinadhuni, B. (2026). Observability and Auditability of Autonomous Agent Decisions in Production Infrastructure. International Journal of Artificial Intelligence and Machine Learning, 6(12s), 1671–1683. Retrieved from https://www.svedbergopen.com/index.php/ijaiml/article/view/3177