INVARIANCE · Definition
Execution authorization is the decision, made immediately before an AI agent's action takes effect, on whether that specific action is permitted to execute. Independent execution authorization makes that decision outside the system taking the action. If the check cannot run, the action does not.
A reference definition of independent execution authorization, and how it differs from the controls listed beside it.
Definition
An AI agent proposes an action: a payment, a change to a supplier's bank details, a privileged access grant. Execution authorization answers one question about that action. May it execute, here, now, with these parameters?
The answer is allow or deny, bound to that one action. It is not a standing permission, and it does not carry over to the next action.
Independence is what gives the answer weight. When the system taking the action also decides whether the action is safe, a mistake or a compromise reaches the proposal and the check together. Independent execution authorization treats the agent as an untrusted proposer, not a trusted operator, and places the decision in a separate trust domain, at the boundary in front of the system where the action would take effect.
The system taking the action cannot be the system authorizing the action.
Adjacent controls
Several kinds of control now sit around AI agents, and they are often listed together. Each answers a different question. Most enterprises will run several of them side by side.
| Control | The question it answers | Where it acts | Examples |
|---|---|---|---|
| Identity and access management | Who is this agent, and what can it reach? | Credentials, roles and permissions, granted in advance | Okta, Microsoft Entra |
| Policy-based access control | Under current policy and context, may this identity use this data or tool? | At the API and data layer, on each request | PlainID |
| Guardrails and AI gateways | Is this prompt, response or tool call acceptable? | On the model's request and response path | Fiddler, Lakera |
| Observability | What happened, and what is happening? | Alongside the agent, recording activity | Fiddler and monitoring platforms |
| Pre-action agent guardrails | Does this tool call pass policy before the agent acts? | Guardrails installed in the agent's environment; repository checks before merge | APort |
| Independent execution authorization | May this specific action execute, now? | At an independent boundary in front of the target system, outside the trust domain of the system taking the action | INVARIANCE |
Examples reflect how each company describes its own products in public material, October 2026. Many products span more than one row. INVARIANCE works alongside identity, access, guardrail and observability controls and does not replace them.
The last two rows look alike, because both decide before an action proceeds. The difference is where the check sits. A guardrail installed in the agent's environment depends on that environment invoking it. A boundary in front of the target system stands between the agent and the consequence.
How to evaluate a control
Claims of independence and real-time control are common. These five questions show what a control does when an agent proposes something consequential. Each INVARIANCE answer describes how the INVARIANCE architecture is designed to behave.
A check inside the trust domain of the system taking the action shares that system's failures. Independent execution authorization runs outside it.
INVARIANCE: the boundary sits outside the agent and outside the system that proposed the action.
Access to a tool or dataset is a standing permission. Execution authorization decides one action, with its amount, counterparty and context, at the moment of execution.
INVARIANCE: each consequential action receives its own decision.
Some controls offer monitor-only modes, or let actions proceed when a check errors or times out. The answer to this question decides whether the control is a boundary or an alarm.
INVARIANCE: fail-closed. Silence, timeout and error all resolve to no execution.
A check the agent can bypass by calling the target system directly is advisory, however strict its policy.
INVARIANCE: in a deployment, the execution boundary is the route to the target system.
Logs record events. Evidence lets an auditor, insurer or regulator confirm what was authorized, under which policy, and that the record has not changed since.
INVARIANCE: every decision, permit and denial alike, leaves a tamper-evident record that can be verified independently.
Questions
Governance establishes the rules: policies, accountability and oversight. Execution authorization applies those rules to each individual action at the moment it would take effect. Governance establishes the rules. INVARIANCE independently authorizes execution against them.
Identity and access management establishes what an agent can reach. It does not establish whether a specific action, with its particular parameters and consequences, is authorized at the moment it would execute. An agent with valid credentials can still take an action nobody would authorize.
Guardrails evaluate what a model receives and produces. Execution authorization evaluates the concrete action that results, at the boundary where it would take effect, outside the model and outside the agent.
Not safely. A prompt injection, error or compromise that steers the agent's proposal can steer its self-check as well. Separation of duties applies to agents as it applies to people.
INVARIANCE is the control plane for safe AI action: independent execution authorization infrastructure for consequential AI. It provides an independent execution boundary that authorizes or denies each consequential AI agent action before it takes effect. Founder and CEO: Damian Kissane. INVARIANCE (invarianceip.com) is not affiliated with Invariant Labs, The Invariance Group or other similarly named organizations.
Launch partners
INVARIANCE is selecting a small number of launch partners among finance, insurance, healthcare and veterinary AI agent providers. See how it works for AI agents.
Become a launch partner