AoA Agent Registry: a dated snapshot of the running agents of the Assessment on the AI framework, 4 September 2026

Abstract

Version 5 (8 September 2026). Twenty agents are documented, up from eighteen: THETA and IOTA were running on the snapshot date, were read in source, and were omitted from every earlier version. REPAIR moves from the unread to the read class; as built it has no machine-solve review and satisfies none of its framework's five invariants. Five claims are narrowed against a second reading of the code (COREA, BASIC, ALPHA, BEST, ADAM), one internal contradiction is corrected (J-CRUD), one claim is measured against the literature (the solver/learner boundary; Berzak, Malmaud & Levy, 2020), and warrant tiers are stated per execution path where an agent has more than one. The registry text records each change where it occurs.This is a dated snapshot, not a maintained catalogue. It records which agents of the Assessment on the AI (AoA) framework were running on the CAFA platform on 4 September 2026, what each was instructed to do, where a human decision was required, and what kind of guarantee each output carries. Twenty agents are documented. Sixteen have been read in source and carry warrant tiers; two, REPAIR and CIA, carry descriptions and no tiers, and say so. Version 1 described every agent from the platform's own documentation, and the documentation was wrong about behaviour more often than not — in one case assigning a warrant tier to a computation the agent does not perform. Version 2 therefore states tiers only where the source has been read. Version 1 also omitted four agents of the same family, which are included here. Four bibliographic errors are corrected, and five claims are narrowed or withdrawn, each at the point where it occurs. The platform is under active development and its agents change. This snapshot is not updated as they do. A running platform closes the question of whether such a system can be built and leaves the question of performance entirely open: no usage data, output quality data, or learner outcomes are reported. The human checkpoints documented are specified rather than measured — each entry states where a human decision is required, not how often it was exercised. Version 4 removes two files that were uploaded to this record in error in version 3; they belong to separate records and are deposited there. The registry text is unchanged from version 3.