Skip to content
IAM Insights

Explainers · Glossary

Action-space

action space

The range of actions an AI agent can take, the tools, systems and permissions it is allowed to use, as distinct from its autonomy (how much it decides for itself); the two together set an agent's risk profile.

In our briefings

The claim each of the newest briefings on Action-space would have you leave with.

  1. Singapore's IMDA offers agent identity practices for organisations to consider in the interim, and names the gaps that make them interim: today's authorisation systems typically use static scopes, and authentication systems, typically built around a single person, struggle when agents act for several users or spin up sub-agents. Each agent should have its own unique, cryptographically verifiable identity, tied to a supervising agent, a human user or a department. The capacity it acts in, such as on its own or for a named user, should be recorded, and every agent identity issued and tracked from a central system. Authorisations should generally be scoped, time- or session-bound, non-transferable and least-privilege, and, as a rule of thumb, a user should not be able to give an agent permissions beyond the user's own. IMDA points to OpenID's and the Cloud Security Alliance's agent identity papers for more.

    Singapore's regulator: each AI agent should have its own identity and, as a rule, no more access than its userBI-FIND-2026-0140

  2. The authors write in their ethics statement that hooking a large language model up to outside environments "has potential dangers, e.g. looking up inappropriate or private information, or taking harmful actions in an environment." Their safeguard was the test design. They kept the web-facing runs to Wikipedia and WebShop, sites they describe as free of private information, and left dangerous actions out of the action space, so a model could not really buy a product on WebShop or edit Wikipedia. The paper's text never uses the words permission, authorization or credential. Its one demonstration of a person correcting the agent, by editing two of its thoughts mid-run, is an ALFWorld example the authors leave for "more systematic study". We read this as safety set by what the action space leaves out: once an action is in it, nothing in the loop the paper describes asks whether the model should take it.

    ReAct's agents searched Wikipedia and a test store; the authors limited risk by offering no dangerous actionsBI-FIND-2022-0007

Cited in 2 briefings

Every briefing that discusses this term, newest first.

  1. Singapore's regulator: each AI agent should have its own identity and, as a rule, no more access than its user

    Article · · Infocomm Media Development Authority (Singapore) · May 20, 2026

  2. ReAct's agents searched Wikipedia and a test store; the authors limited risk by offering no dangerous actions

    Paper · · arXiv · Shunyu Yao et al. · Oct 6, 2022