Abstract

Abstract (Part 18 of 27)

Agent A sends an authenticated instruction to resource domain B to transfer 150 units under an exact idempotency identity. The resource applies the operation, commits its local ledger, and transmits a completion receipt. The response is lost during a network partition. Agent A’s supervisor sees only a timeout and proposes to retry. Meanwhile the original coordinator restarts and a new instance is migrated to another host. Both can still possess historically valid signatures. If either repeats the transfer, if the old instance retains effective authority, or if the destination uses an incomplete checkpoint, the external system may suffer a duplicate effect or unsafe next-phase advancement. The essential difference is between communication failure and effect failure. It is not safe to treat lack of response as permission to repeat an action. A newly elected or migrated coordinator does not obtain native resource authority merely because its scheduling state is current.

Difference from earlier parts: Earlier embodiments control the joint plan, the freeze, evaluation, next-actor selection, role reconfiguration, selective communication, the next-move passport, Finality Sink enforcement, substitute selection, pending-work escrow, verified phase advancement, actual-outcome feedback, and independent evaluator integrity. This embodiment specifically controls cross-domain network partitions, unknown outcomes, durable stale-authority fencing, checkpoint-bound migration, and independently authorized distributed recovery — determining whether a resource actually acted when communication and control state diverged, and which entity currently retains effective authority.

This embodiment (Part 18 of a 27-part series under Indian Patent Application Number 202631131650; full details at https://zenodo.org/records/23940766) establishes resource-scoped original-operation identifiers, native-effect hashes, idempotency keys, current protected generations, and a journal of authoritative outcomes. On detection of a material partition or authority conflict, a protected controller places applicable effect paths in HOLD, advances a durable fencing generation, and directs read-only status inquiries to native effect-owning resources. Historical committed operations are distinguished from absent answers, unconfirmed operations, disputed results, and states that cannot be resolved. A new coordinator or destination workload is admitted only under matching authenticated migration lineage, current resource policy, and separately verified authorization. Receiver-side checks reject obsolete leaders, repeated operations, inappropriate destinations, unauthorized phase changes, and stale approvals. Where operations span resources without a common atomic transaction, the method uses explicitly classified partial effects and separately authorized recovery or compensation, rather than assuming a global rollback. The architecture prioritizes preventing false-positive effectuation over mandatory progress.

The architecture therefore contains the damage from partitions, lost responses, restarts, and migrations in multi-agent systems without treating silence as success or allowing stale authority to continue.

Keywords: multi-agent systems, agentic AI, AI safety, AI governance, network partition containment, unknown outcome reconciliation, durable fencing, stale-authority fencing, checkpoint-bound migration, fail-closed systems, zero-trust agents, idempotency, authoritative status verification, cross-organizational collaboration, execution finality, communication failure vs effect failure, read-only status inquiry, protected recovery, independently authorized distributed recovery, resource-side epoch enforcement, non-assumption of noncommit, partial-effect tracking.

Creative Commons License

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

Download

Share

COinS