Year: 2026
Type: article
Abstract: An enterprise agent may produce a valid tool call while the connected system rejects the request, accepts it for later processing, or changes a record without returning confirmation. The action connector layer must preserve these distinctions. This article develops a design model for that layer by analyzing recent research on Application Programming Interface (API) use, stateful tool execution, and agent security. It specifies an action contract with typed inputs, preconditions, an authority scope, an expected effect, and defined outcomes. The model classifies integrations by credential ownership and the extent of control over the service interface. It then relates validation errors, authorization failures, provider rejections, and uncertain execution states to permissible agent responses. A versioned contract provides multiple agents with a common definition of an action, while the connector enforces permissions and interprets the external response. The proposed specifications support connector review, provider changes, and the diagnosis of failures across enterprise workflows.
Author: Garg , Nitin
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7233534976
