Understand local agent context.
Review supported local AI agents and MCP context alongside covered AI traffic on managed devices.
Local agent visibility is available in selected rollouts. Inventory does not establish control over every tool action.
Workforce AI SecurityAI agent guardrails.
AI agent guardrails.
Check before
the next call.
Check content at an agent checkpoint you control, then enforce the result in your application.
You choose the checkpoint. Your integration enforces.
01 / Your app sends text
Customer reply with account detail
Draft a reply to alex.morgan@example.invalid about account SAMPLE-104.
02 / Aona content checkReject
The example policy flags customer information in this text.
03 / Your application acts
Stop the model call.Your application keeps this text from the model. It can use a validated redacted version if configured.
Policy violation selected. Aona check: reject. Your application: Stop the model call.
Where Aona fits
Two routes. Different responsibilities.
Choose visibility on a managed device, evaluation inside your application, or both where supported.
Review endpoint coverage
Check content at your checkpoint.
Submit text to a framework or guardrail evaluation, or supported files to a framework evaluation.
Your application applies the verdict, returned redaction and failure policy. Calls outside the integration are outside that checkpoint.
Illustrative application reviewBefore an external action
Draft reply to [CUSTOMER] about account [ACCOUNT].
- Target
- Ticket SAMPLE-104
- Action
- Send drafted response
- Evidence
- Source ticket and draft
A reviewer makes the decision in your application.
Build around the checkpoint
Review the action before it happens.
Give the reviewer the target, proposed change and evidence for consequential decisions.
- Limit access.
- Review the action.
- Verify the result.
Your application owns permissions, approval and execution.
Designing a meaningful review
Limit permissions
Give the agent only the tools, identity and access needed for its task. Enforce those limits in trusted systems.
Make review meaningful
For consequential actions, show a person the proposed target, action and evidence before approval.
Verify the result
Check generated outputs and actual system state before treating the task as complete. Preserve useful evidence.
Evaluation and evidence
Know what was checked. Record what happened.
A verdict describes the evaluation. It does not prove that a request was stopped, a tool was restricted or a person approved an action.
Stateless evaluations do not automatically create analytics events. Record the action your integration took, and keep evaluation APIs distinct from APIs that retrieve existing events.
Understand event APIs and webhooksTake it into your evaluation
Know what you are evaluating.
Use these questions with your security team or an AI vendor. A “yes” on a feature list is only the start.
- What is covered?
- Who enforces the decision?
- What evidence records the result?
Open all seven evaluation questions
- 01
Which user, app, agent and input path does this control cover?
- 02
What risk and policy does the check address?
- 03
What happens when the check allows, rejects, times out or fails?
- 04
Who applies the decision, and can another path bypass it?
- 05
Where is data processed, stored and retained?
- 06
What evidence records the action actually taken?
- 07
Which allowed, prohibited and ambiguous examples will we test?
Questions
Your questions, answered.
What are AI agent guardrails?
Agent guardrails are controls around an AI agent's inputs, permissions, tools, actions and outputs. Examples include evaluating sensitive data before a model call, limiting tool access and requiring human approval for consequential actions. Different layers address different risks; a content check alone cannot constrain everything an agent can do.
How can we use Aona with an AI agent?
Your integration can call Aona framework or guardrail APIs to evaluate supplied text; framework APIs also evaluate supported files. You choose the checkpoint and apply the returned verdict or redaction. Define timeout and error handling, cover alternative paths and record the decision your application actually took. Aona endpoint products separately provide supported AI traffic protection and local agent visibility in selected rollouts.
Will the API stop an agent by itself?
No. An evaluation API returns a result to the caller. Your application must enforce the decision before continuing. Calls that bypass the integration are not covered. The event API and outbound webhooks deliver existing event evidence; they do not themselves intercept an agent action. Stateless policy evaluation does not automatically create an analytics event.
Does local agent visibility mean every tool action is controlled?
No. Seeing an agent, MCP configuration or local tool context is different from enforcing every action it can take. Local agent visibility is available in selected rollouts. Validate the specific application, request path and response you need rather than inferring action control from inventory.
Where should human approval fit?
Use human approval where the consequences justify it, such as sending an external message, changing production data or committing funds. Give the reviewer the proposed action, its target and relevant context. This is an application design practice; Aona's evaluation APIs do not by themselves create approval queues, grant permissions or guarantee that a person reviewed the action.
Aona AI guardrails
Bring your agent checkpoint.
Walk through the content, evaluation and response your integration needs.