The words, in plain language
This category has a lot of vocabulary and not much agreement on what any of it means. These are the terms we use, defined the way we would explain them to a colleague.
Terms
Each definition leads with a direct answer, then says what it means in practice.
A
- Accountability
- Being able to say what an agent did, on whose authority, and who answers for it. It is a stronger claim than monitoring. A log tells you what happened; accountability means the action was attributable to a named owner before it was taken.
- Agent discovery
- Finding the agents already running in an organisation, including the ones nobody registered. Most companies have more than they expect, because building one no longer requires engineering. Discovery is the starting point rather than the goal: an inventory is only useful once you can act on it.
C
- Control plane
- The single place where discovery, identity, policy and intent are managed together. The alternative, one tool per question, is how teams end up able to answer what agents exist but not what any of them is allowed to do.
D
- Delegation
- Handing an agent the authority to act for you, within limits you set. The useful comparison is a colleague rather than a tool: you would say what they may decide alone, what needs checking first, and what is not theirs to touch.
E
- Enterprise AI agent
- Software that takes actions inside a company on behalf of a person or a team, rather than only answering questions. The distinction that matters is not how clever it is. It is that an agent can log into systems, move data and make changes, which means it needs the same joining, permissions and oversight as anyone else who can do those things.
H
- Hybrid identity
- The pairing of an agent with the person or team accountable for it. An agent acting alone is difficult to reason about. An agent carrying both its own identity and the identity of whoever stands behind it can be reviewed, scoped and, when the person moves on, dealt with.
I
- Inherited permissions
- An agent taking its access from the person who runs it, rather than being granted its own. Inheritance is the honest starting point, because it means an agent can never quietly hold more than its owner does. Scoping down from there is the work.
- Intent
- What an agent is about to try to do, checked before it happens rather than recorded after. Reading intent at the moment of the request is what makes it possible to stop a mistake before data moves, instead of finding it later in a log.
J
- Joiner, mover, leaver
- The three moments in any working relationship where access should change: someone arrives, someone changes role, someone leaves. Agents have all three. The leaver step is the one most often skipped, which is how an organisation ends up with working credentials attached to a project that ended a year ago.
L
- Least privilege
- Giving an agent exactly the access its job needs, and nothing beyond it. The practical version is narrower than the principle sounds: not read access to a whole system, but the specific records the task touches, for as long as the task runs.
O
- Onboarding an agent
- Giving a new agent an identity, an owner, a scope and a review date before it starts work. The phrase is deliberate. An agent that arrives through the same door as a new colleague inherits the same clarity about what it is for and who is responsible for it.
P
- Provenance
- The record of where an action came from: which agent, acting for whom, under which policy, using which data. Provenance is what turns an audit from an archaeology exercise into a question with an answer.
S
- Scoped down
- Reducing the access an agent inherited, to the part it actually needs. A person with broad permissions who builds an agent for one narrow task should not hand that agent the whole of their access, and scoping down is how the gap is closed.
- Standing access
- Permissions that remain in place whether or not they are being used. Standing access is convenient and it accumulates. The alternative is access granted for a task and withdrawn when the task finishes, which keeps the total amount of open permission close to what is genuinely in use.