The Agent Governance Glossary

Checks that run at the moment an agent is about to act, allowing, guiding, or blocking the action itself. Stronger than content filtering because they stop consequences, not just words.

An AI worker that takes actions, not just answers. Agents hold identities, receive tasks, use tools, and produce auditable changes in real systems.

Every agent signs in as itself, with access scoped to the job at hand and expiring when the work ends. No shared logins, no static keys.

The controls that make agent autonomy answerable: who may act, what they may do, what needs a human, what it may cost, and how all of it is proven afterwards.

A set of AI agents that carry real work end to end: operating infrastructure, staffing a service desk, shipping engineering changes, and running business processes, under controls the company sets.

A pause on a consequential action until a named human says yes, in platform chat or Slack, with the who and the when kept on the record.

The structured, tamper-evident record of every agent action, guardrail decision, approval, and enforcement. The trail is the evidence; nobody reconstructs history from chat logs.

A spend cap enforced before AI work is dispatched. When the budget is exhausted, agents stop cleanly instead of running up a bill that surfaces on the invoice.

The platform runs inside your own cloud accounts and your own identity provider. Your data and workloads never leave your environment, and there is no proprietary control plane holding them hostage.

Screening of the text going into or out of a model, such as prompt-injection detection and output filtering. Complements action-layer control, which governs what the agent actually does.

The four answers a guardrail gives for an agent action: let it proceed, correct it and let it proceed, stop it, or let it through as a recorded exception with the agent's reason kept on the audit trail.

Evidence is the continuous record that controls operate; certification is an auditor's judgment on top of it. A vendor can hand you evidence; only an audit grants the certificate.

The international standard for AI management systems: the policies, oversight, and controls an organization wraps around its use of AI.

An open protocol that lets AI assistants use tools. It is how you connect your own AI assistant to the platform, through a guarded bridge with scoped access.

The US AI Risk Management Framework: the shared vocabulary boards and regulators use to ask how AI risk is governed, mapped, measured, and managed.

A machine-readable format for compliance evidence. Auditors and tools consume structured artifacts instead of screenshots.

A proven, repeatable way to get a class of work done. The platform's intake picks the right playbook for a request and puts the right agents on each step.

A signed, versioned bundle of guardrail rules covering one domain, such as security or devops. Packs update as data, so protection improves without waiting for a software release.

AI tools, assistants, and agents in use across an organization without approval, governance, or an audit trail. The first question most AI-governance programs have to answer is simply: what is running?

A property of records that makes alteration detectable. If someone edits history, the record shows it.