On this page
Stopping a process is not the same as retiring its authority. Root-scoped quiescence accounts for every registered path that can outlive a long-running agent.
Research status: Preprint. Original research date: 2026-09-18.
Authors: Genliang Zhu, Chu Wang.
Overview
Long-running agents leave authority in more places than the process that started them: credentials, delegated tasks, queues, callbacks, reservations, and provider-side operations may continue after cancellation or credential revocation. Some shared work may also have valid support independent of the retired root.
The protocol creates a linearized root cut, fences further old-root expansion, and collects provider-frontier evidence for every registered path. Old-root effects are stopped, while work with a current and independently sufficient authorization can be rebound exactly and continue.
Core contributions
Root-scoped quiescence
Defines revocation as a manifest-bound certificate over every cut-relevant acceptance, rather than as process cancellation or a claim of global idleness.
Delegation-aware authority
Represents alternative and conjunctive support as antichains of minimal sufficient root sets so shared work can be retired or preserved without conflating its authority.
Compositional evidence
Composes provider-frontier certificates into a cutset and uses exact channel-token accounting to reconcile asynchronous transfers, treating missing or conflicting evidence as indeterminate.
Executable revocation checks
Exercises late effects, cuts, fences, restarts, stale processes, replay behavior, and semantic regressions through registered traces and an independently implemented checker.
Evidence and scope
This is an arXiv preprint. Its certificate establishes root-relative authorization quiescence only within the bound manifest and configuration; it does not prove global idleness, rollback, or business completion. The guarantees assume registered old-root paths, exact channel conservation, sound provider-frontier evidence, and the stated fencing and linearization conditions. The reported evaluation is provider-free and trace-based rather than a production deployment study.
