Runtime governance vs. runtime enforcement

What can actually stop an AI agent action?

Runtime governance cannot. It evaluates each action against the scope and policy declared for the run, records what happened and flags deviations, and the action proceeds. Pausing, redirecting or stopping an action takes runtime enforcement: a control point in the execution path that acts on the evaluation.

Most teams start with governance to see what their agents actually do, then add enforcement where a deviation is too costly to discover after the fact.

Three layers, three jobs

Governance evaluates. Enforcement acts. Secure Runtime contains.

Runtime Governance

Sentience Governor · open source

What it does
Evaluates each action against declared scope and policy. Records what the agent did and flags deviations.
Can it stop an action?
NoAdvisory, and fails open. The action proceeds and is flagged.
Declared objective
Recorded with the run as context. Objective-aware evaluation comes with Runtime Enforcement.
Where it runs
Where your agent runs. Local traces, no account, no telemetry.
How it’s offered
Free and open source (Apache 2.0).

Runtime Enforcement

Sentience · custom to your environment

What it does
Acts on each evaluation before the action takes effect: course-corrects, escalates to a person, or stops execution.
Can it stop an action?
YesAt the execution control points integrated with your environment.
Declared objective
Evaluates actions against the declared objective, scope and policy, configured for your workflows.
Where it runs
Your existing infrastructure.
How it’s offered
Built for your environment and priced per governed workflow. Includes Runtime Governance.

Secure Runtime

Sentience · isolated execution

What it does
Runs agents in isolated execution environments, with governance and enforcement integrated.
Can it stop an action?
YesThrough integrated enforcement, scoped credentials and policy-governed egress.
Declared objective
Objective-aware evaluation, delivered through customer-specific integration.
Where it runs
Sentience’s isolated execution environments. Your models and systems connect to it.
How it’s offered
Commercial. Starts with a proof of concept on one workflow.

One action, three outcomes

An invoice agent emails a customer.

The agent is declared for one job: reconcile the August invoices, with scope crm and billing. Partway through the run it calls a tool that sends an email.

Attempted actionemail_customer→ targetemailnot in declared scope (crm, billing)
  1. Runtime GovernanceReal run

    Recorded and flagged.

    SCOPE_INTENT_MISMATCHPOL-001

    Governor records the call before it is dispatched and flags it as outside the declared scope. The email is sent. A reviewer sees exactly what happened, after it happened.

    EmailSent

    From a real run of pydantic-ai-governor 0.1.1.

  2. Runtime EnforcementIllustrative

    Redirected, paused or stopped.

    RedirectedPausedStopped

    With a control point on outbound email, the same evaluation can redirect the message to internal review as a draft, hold it until a person approves, or end the run, depending on policy and risk.

    EmailHeld, redirected or not sent

    Illustrative. Available interventions depend on the control points integrated with your environment.

  3. Secure RuntimeIllustrative

    Governed path, enforcement built in.

    scoped credentialspolicy-governed egress

    The agent runs in an isolated environment, so outbound email is a governed path rather than an open one, and enforcement is already integrated.

    EmailOnly through governed egress

    Illustrative.

Choosing

Start by seeing. Add control where it counts.

01 · See

Start with Runtime Governance

When you need to see what your agents actually do, keep a record for review or audit, and learn which deviations matter. It is free, and it records without intervening in the run.

Runtime Governance

02 · Control

Add Runtime Enforcement

When some deviations are too consequential to find after the fact: payments, deletions, external communication, production changes. It acts where your agents already run.

Runtime Enforcement

03 · Contain

Choose Secure Runtime

When agents also need isolated execution infrastructure: scoped credentials, controlled network access, and a path from a validated experiment to a production template.

Secure Runtime

Evidence boundaries

What governance alone cannot do.

  1. 01It cannot stop, pause or alter an action, and must not be relied on as a control.
  2. 02It flags actions outside the declared scope and policy. It cannot determine that an in-scope action serves an objective nobody approved.
  3. 03Recording a task’s declared purpose is not the same as authorizing it.
  4. 04It says nothing about runs it did not observe.

Questions

Common questions.

Can Sentience Governor block an AI agent action?

No. Sentience Governor is advisory. It evaluates each action against the declared scope and policy, records it and flags deviations, and the action proceeds. It fails open and must not be relied on as a control. Pausing, redirecting or stopping actions is Sentience Runtime Enforcement.

Does Runtime Enforcement require Secure Runtime?

No. Runtime Enforcement works through control points integrated with your existing infrastructure, and it includes Runtime Governance. Secure Runtime is for teams that also need isolated execution infrastructure.

Do I have to replace my agent framework?

No. Sentience Governor works with Claude Code, Codex and other coding agents, Pydantic AI, MCP clients, LangChain and LangGraph. Runtime Enforcement works where your agents already run. Secure Runtime connects to your existing models, frameworks, gateways and enterprise systems.

What is an execution control point?

A place in your agent’s execution path, integrated with your environment, where a decision can take effect before an action does. Which interventions are available depends on which control points are integrated.

Next step

See what your agents do first. Then decide what to stop.

Start with Sentience Governor

Already running agents on Pydantic AI? Follow the five steps