Abstract

Abstract

An AI agent pays a vendor, but the invoice is never booked and finance is never told. Every individual step was authorised and no credential was forged, yet the world that occurred is not the act that was approved. Many actions take effect only as several consequences at several systems: a settlement post, a ledger write, a tool call, a notification, a device enable. When one intended action works this way, correct authorisation at each system does not make the whole act final.

This disclosure describes all-or-none execution authority for such composite acts. One composite act is described by a canonical record listing every child action, its target system (Finality Sink), and whether it is required. A composite act digest is computed over the sorted child list. A protected enforcement domain issues one non-bearer, single-use child execution handle per required child. Each handle is bound to the child’s own act digest, its sink, the composite identifier and the composite digest, so altering the child list invalidates every handle.

Each sink reconstructs its own live pending effect, verifies its handle, and prepares by reserving consume-state without any external effect. A coordinator may only prepare and drive. It must first win a write to a durable, linearizable Composite Decision Log, which holds one value per composite (empty, COMMIT or ABORT) and accepts only compare-and-swap writes, so the first writer wins. The successful COMMIT write is the finality point.

A prepared child whose reservation timer expires consults the log. It commits on COMMIT, releases on ABORT, writes ABORT itself if the log is empty, and enters a no-effect HOLD only if the log is unreachable, so no child ever acts on silence. A safety theorem shows that no execution can contain both a posted and a released required child, without relying on the coordinator being correct or alive.

If a required child cannot prepare, the honest claim is mitigation, not prevention. Any reversal is a new compensation act with its own handle, never a silent undo of a consumed handle. The approach complements two-phase commit, XA, sagas, TCC and workflow engines rather than replacing them. It applies to AI agent tool chains (including MCP), payments (authorise and capture), enterprise ERP, control planes and cyber-physical systems. It supports EU AI Act record-keeping, human oversight and robustness duties, and financial-sector resilience rules such as DORA.

This technical disclosure is based on the Mothership application PCT/IB2026/055615, filed 4 June 2026 and publicly disclosed 16 July 2026 (WO 2026/150382). That application claims priority from PCT/IB2026/053385, filed 7 April 2026 and publicly disclosed 4 June 2026 (WO 2026/115520). The protocol details are as published in Internet-Draft draft-das-composite-execution-finality-01 on 16 September 2026. Related family applications can be cross-checked on WIPO PATENTSCOPE.

The industry-relevance material is AI-generated, and readers must cross-check it against the original PCT filings before drawing any conclusion. Licensed under CC BY-NC 4.0. The author is open to acquihire or strategic collaboration.

Keywords

composite execution finality; all-or-none execution; partial commit prevention; split effectuation; multi-sink authorization; AI agent tool chain atomicity; MCP tool-call authorization; agentic AI safety; two-phase commit; XA; saga pattern; TCC; compensating transaction; compare-and-swap decision log; linearizable register; first-writer-wins; coordinator failure; non-blocking commit; reservation timeout; authorise and capture; payment orchestration; execution handle; non-bearer capability; atomic consume; finality receipt; act digest; fail-closed; EU AI Act; DORA; machine enforcement.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License

Download

Share

COinS