ORIGINAL TECHNICAL PERSPECTIVE
How should enterprises govern AI agents that can take action?
Separate model suggestions from execution authority. Make identities, least privilege, approval boundaries, and auditable tool actions explicit.
THE SHORT ANSWER
Evidence before confidence.
Govern the authority to act, not just the words an agent generates. Define an accountable owner, the identity used for each resource, the operations it may perform, and the actions requiring human approval. Enforce permissions in the tool or service that executes the operation. Evaluate behavior across tool calls and tool responses as well as the final answer, and preserve enough evidence to investigate and stop an unsafe workflow.
Architectural judgment
Where should the trust boundary sit?
A helpful personality and a careful system prompt do not establish permission to change a business system. The model may propose an action; a trusted execution layer must decide whether the authenticated actor is allowed to perform it, on that resource, under those conditions. Treat retrieved documents and tool responses as information, not a source of new authority.
Start by separating read-only assistance from actions that send messages, modify records, grant access, or create external commitments. Introduce new authority deliberately. Each additional tool should have a clear purpose, a restricted contract, observable failure behavior, and an owner who can disable it.
Supporting reading: Microsoft Learn: Agent identity concepts; Microsoft Learn: Guardrails and controls.
Microsoft Learn guidance
Which identity is acting, and with what permission?
Microsoft Learn describes Foundry agent identities as a way to authenticate agents to downstream resources and govern their access. The documentation distinguishes identity and token issuance from authorization at the target resource, where role assignments or another supported permission model determine access. An identity existing in a directory is not itself approval for an operation.
Map the user, agent, and downstream resource explicitly. Check the authentication mode supported by each tool connection instead of assuming every call runs with the user's permissions. Apply the narrowest appropriate roles and scopes, and test denied access as deliberately as successful access. Do not put secrets or authorization rules inside model instructions.
Supporting reading: Microsoft Learn: Agent identity concepts.
Architectural judgment
When should a person approve an action?
For consequential actions, the approval should describe the specific operation, target, and material inputs. A vague instruction to be cautious is not an approval workflow. Validate the request before execution and ensure that the action being executed is the action that was approved, rather than allowing a changed tool call to reuse an earlier decision.
Choose the boundary according to consequence and reversibility. Reading an allowed document and sending a message to a customer do not need identical controls. This is a design recommendation, not a claim that every Foundry tool provides a ready-made approval mechanism. Where the application needs that guarantee, implement and test the enforcement outside the model.
Supporting reading: Microsoft Learn: Agent identity concepts; Microsoft Learn: Guardrails and controls.
Microsoft Learn guidance
Do platform guardrails replace application controls?
No single platform feature establishes end-to-end safety. The Foundry guardrails documentation checked for this perspective marks agent guardrails, tool-call and tool-response intervention points, and hosted-agent network egress controls as preview. It also limits the documented agent guardrail system to agents developed in Foundry Agent Service, not every agent registered in the control plane.
Verify support for the actual agent type, model, region, and API version before designing around a control. Filtering behavior depends on configuration and application design. Continue to enforce authorization, validate tool inputs, review data shared with external systems, and test the full workflow. A content-safety result is not a substitute for permission to act.
Supporting reading: Microsoft Learn: Guardrails and controls; Microsoft Learn: Agent identity concepts.
Decision checklist
What should a governance review ask to see?
Review the trace of an action, not only the final conversational answer. An agent can produce a reasonable-sounding response after an inappropriate tool call. Conversely, a denied operation can be the correct outcome. The evidence should make that distinction visible while limiting access to sensitive inputs and outputs.
- The accountable owner, intended task, allowed resources, and explicit exclusions.
- The agent and user identity model, scoped permissions, and negative authorization tests.
- The approval contract for consequential operations and evidence that it is enforced at execution.
- Tests using hostile instructions in user input and tool responses, with documented residual risks.
- A privacy-aware audit trail, operational alerts, and a procedure to revoke access or stop the agent.
Supporting reading: Microsoft Learn: Agent identity concepts; Microsoft Learn: Guardrails and controls; Microsoft Learn: Observability in generative AI.
Sources, scope, and currency.
Original technical synthesis, not a customer deployment report. Recommendations are distinguished from documented platform capabilities. No customer results, adoption figures, or universal performance thresholds are claimed.
Microsoft Learn sources checked on . This is a dated review, not a live documentation feed or a claim that every capability is generally available. Recheck the current documentation, preview status, and deployment requirements before implementation.