Action-space is the range of actions an agent can take, including the transactions it can execute, determined by the tools it is allowed to use and its permissions on those tools. The term is also given as authority or capabilities. Model AI Governance Framework for Agentic AI presents it as one of two concepts worth distinguishing when asking what an agent can do, the other being autonomy: the degree to which an agent can decide how to act towards a goal, such as by defining the steps to be taken, determined by its instructions and the level of human involvement.
Usage
An agent’s action-space depends mainly on the systems its tools reach — sandboxed tools that cannot affect anything else, internal systems such as the organisation’s own databases, or external systems reached through third-party APIs — and on whether it can only read from those systems or also write to and modify them. A computer use agent is described as an emerging modality that expands the action-space sharply, because access to a computer and browser lets it take any action a human could take with them (scrolling, clicking, typing) without relying on specifically defined tools and APIs.
Autonomy, the other axis, turns on how prescriptive the instructions are — a detailed SOP versus an instruction to use the agent’s own judgment — and on where the deployment sits on a spectrum of human involvement running from agent proposes, human operates through agent and human collaborate and agent operates, human approves to agent operates, human observes. The two axes are treated as independent: the framework illustrates them with a software engineering agent that can search and read any file to diagnose issues but has no write access (more autonomy, smaller action-space) set against one that can browse the web and modify files but only with human approval at each step or under a fixed SOP (less autonomy, larger action-space).
Separating the axes matters for risk assessment, where they enter on different sides: the scope of an agent’s actions and their reversibility are listed among the factors affecting the impact of a failure, while the agent’s level of autonomy is listed among the factors affecting its likelihood, on the stated reasoning that higher autonomy produces more unpredictability.
Related Terms
Agentic Autonomy Levels, Supervised Agency Spectrum, Human-in-the-loop, Computer Use, Sandboxing