Abstract

This disclosure describes how AI agents behave when a service they depend on, such as an email or calendar service, a payment provider, a data source or another AI model, fails or becomes unreliable. Each certified agent declares its dependencies in a public registry, with the criticality of each and the fallback routes approved in advance for it, such as another certified provider, a sealed envelope over a different transport, or handover to a person; fallback routes are rehearsed on the agent's certified twin. When a dependency fails, as judged by a detection framework that distinguishes an unavailable service from an overloaded one or from the agent's own fault, the agent declares its degraded state to the person it acts for: what it cannot do, what it has queued, and what it has dropped, and it never reports a task as done when it is not. Retries are bounded, spaced and randomized, and when many agents depend on the same failing service, a coordination layer that can only tighten reduces their combined retry budget, so that agents do not add to the overload. Actions that cannot be completed are queued with an expiry and a key that prevents duplicate execution on recovery; an action whose circumstances may have changed while queued, beyond a declared staleness, is confirmed again with the person before it runs, and a queued payment is checked again against the agent's spending limits at the moment it executes. An agent switches only to fallback routes declared in advance and of equal or stronger assurance, records each use, and never improvises another channel, account or method to complete a task. If a dependency needed for the agent's own oversight fails, the agent tightens its limits instead. When the service returns, queued work resumes in stages with load ramped, and the person receives a reconciliation of what was done, delayed, dropped or prevented from duplicating.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Share

COinS