Agentic AI Governance Glossary

AI systems that pursue a goal over multiple steps and take actions in the world — calling tools, reading and writing data, and invoking other systems — rather than only returning text to a person.

At mAIndala: mAIndala treats the action, not the answer, as the thing that needs governing: what an agent is permitted to touch, and what record it leaves behind.

A program that uses a language model to decide what to do next, equipped with tools it can call and a goal it is working toward. Unlike a chatbot, it acts; unlike a script, its sequence of steps is not fixed in advance.

An open standard for connecting AI models to external tools and data sources through a uniform interface, so a capability written once can be used by any compatible client.

At mAIndala: mAIndala maintains a public catalog of MCP servers and can govern the calls an agent makes through them.

See it in context →

A single layer that governs every AI agent an organization runs, regardless of where each agent was built: what exists, what each one may do, which credentials it may hold, and what record its actions leave. "Control plane" borrows the networking sense — the layer that decides and enforces, kept separate from the systems doing the work.

At mAIndala: This is the category mAIndala builds in, which is why governance is applied at boundaries the platform owns rather than by asking each agent to police itself.

See it in context →

AI tools and agents in use inside an organization without the knowledge or approval of the people accountable for security, privacy, or compliance. The agentic version is materially riskier than shadow SaaS, because an unregistered agent holds credentials and takes actions rather than merely storing data.

A record of the AI agents actually running in an organization, including the ones nobody registered. It is the agentic equivalent of an asset inventory: you cannot govern, risk-rate, or decommission what you cannot enumerate.

At mAIndala: mAIndala can build this inventory from scans an organization chooses to run and upload, and keeps discovered agents alongside registered ones. It is opt-in and metadata-only — mAIndala does not continuously monitor a customer network.

See it in context →

An itemized record of what an AI agent is made of — the models it calls, the tools and data sources it reaches, and the other agents it delegates to. The idea is borrowed from the software bill of materials (SBOM), applied to systems whose composition changes as they run.

See it in context →

A control point that sits between an agent and the MCP tools it calls, applying an organization’s policy to each call before it reaches the tool — and recording what happened. Gateways of this kind exist because the alternative, trusting each agent to enforce its own rules, gives an organization nothing it can audit.

At mAIndala: mAIndala operates a policy-controlled tool gateway with allow and deny rules, rate limits, time-of-day windows, and DLP redaction, plus a full audit trail of governed calls.

See it in context →

Checks applied to an agent’s activity while it is running, rather than only when it is built or reviewed — for example inspecting a tool call for prompt injection, attempted data exfiltration, or credentials in transit, and blocking or flagging it before it proceeds.

At mAIndala: mAIndala applies runtime guardrails to the tool calls that pass through its gateway. They govern what an organization routes through that boundary; they are not a claim to see inside an agent’s own reasoning.

See it in context →

A rule set naming which tools an agent may and may not call. The agentic counterpart of least privilege: an agent that can reach every tool available to it has the blast radius of every tool available to it.

See it in context →

A choice about what should happen when a governance check cannot be completed. Fail-closed denies the action; fail-open permits it. The trade-off is availability against assurance, and which one is correct depends on what the agent is about to do.

At mAIndala: Organizations choose this for themselves. By default a governed call proceeds when the policy check cannot be completed; an organization that would rather stop than risk an ungoverned call can turn fail-closed on.

See it in context →

A control that immediately stops an agent from acting — cutting off its tool access and credentials — without needing to redeploy it or reach whoever built it.

At mAIndala: mAIndala provides an instant kill-switch for governed agents, including agents built on other platforms.

See it in context →

Encrypted storage that holds the secrets an agent needs and issues scoped, short-lived credentials at the moment of use, so a long-lived key is never embedded in the agent itself. Revoking access then becomes a decision, rather than a hunt through deployments.

At mAIndala: mAIndala issues scoped, short-lived credentials to governed agents from an encrypted vault.

See it in context →

One agent handing work to another. It is what makes multi-agent systems useful and what makes them hard to govern: without a control on delegation, an agent can reach capabilities it was never granted directly, by asking a peer that holds them.

At mAIndala: When one agent delegates to another inside an organization, that delegation is checked against the organization’s policy and recorded.

See it in context →

A cryptographically checkable answer to "which agent is this?", so a system receiving a request can confirm the caller is the agent it claims to be rather than something impersonating it — and so that access can be revoked for one agent without affecting the rest.

At mAIndala: Agents governed by an organization can carry a verifiable identity. Identities can be revoked, which stops anything presenting that identity from being accepted, typically within seconds.

See it in context →

A time-bounded, tamper-evident export of the records showing what a system was permitted to do and what it actually did, in a form an auditor or regulator can check independently. The distinction that matters is between a report assembled about a system and a record generated by the system as it ran.

At mAIndala: mAIndala exports signed, timestamped evidence packs that a third party can verify offline, without a mAIndala account. They are a record of what happened — not a certification, and not a claim of compliance.

See it in context →

A published rating of how much assurance a catalog entry carries — how it authenticates, what its documentation and provenance look like, and what an automated security review found — so a team can compare options on more than popularity.

At mAIndala: mAIndala rates catalog listings with an automated security review and publishes the resulting score and badge on each listing. A score reflects the review that produced it, not a live reading.

See it in context →

A durable, ordered record of actions taken and decisions made, kept so that a later reader can reconstruct what happened without relying on anyone’s recollection. For agentic systems it is the difference between believing an agent behaved and being able to show it.

At mAIndala: Every governed call is recorded with the agent behind it and which control allowed or blocked it.

See it in context →