Agent Record Test: Method, Version 0.1
Year: 2026
Type: article
Abstract: This working paper sets out a method, fixed and deposited before any testing, for finding out whether an ordinary user can obtain the record of an act done in the user’s name by an AI agent. Ten publicly offered AI agent products each perform one small task (sending one message to a test team address, or making one cancellable booking of not more than INR 500), twice, by two testers on their own accounts. Each tester then looks for the record in the product and asks the provider for it in writing through its published channel. What is obtained within a response window of seven days, equal for every run, is scored against the nine field minimum agent record proposed in clause 2.4 of Addendum Version 1.2 to the Duggal Global Agentic AI Liability Framework: authorisation and scope, instruction chain, timestamps, model and version, tool calls, approvals, incident records, integrity mark, and retention and custodian. The method fixes in advance the sample selection rules, the task, the request wording, the scoring rules, the headline figure and the further figures that must be published with it, the right of reply, and the publication rules. The list of ten products is deposited as a dated supplement, linked to this record, before the first test is run. No product has been tested as at the date of this version, and this version names no product and no company. This is an exploratory pilot run from India and a test of transparency to the user. It is not a legal compliance audit, a security audit, a product ranking or a finding of liability. The nine field record is the author’s own proposal and is not a legal requirement in any jurisdiction. A field scored “Not obtained” means the tester could not obtain it, not that the provider does not keep it. This version has not been externally reviewed. Comment on the method is invited and will be answered in a later version.
Author: DUGGAL, PAVAN
Language: English
Subfield: Law
Field: Social Sciences
Domain: Social Sciences
Keywords: agentic AI, records retention
Open Access status: green
OpenAlex ID: W7225861876
