On 18 November 2026, Microsoft announced Agent Controls, a managed Azure service available in public preview. It allows organisations to evaluate policies governing AI agents while they operate. This is a development leaders should examine carefully: governance needs to reach the point at which an AI system can take action, while there is still time to intervene.[1]
Runtime simply means while the system is running. A policy approved before deployment and an audit conducted afterwards serve different purposes from a control that can stop a proposed action before it happens. Organisations deploying agents need to understand how these arrangements work together.
Microsoft describes intervention points across an agent’s lifecycle, including model calls, tool calls and outputs. Evaluations can combine fixed rules, classifiers and AI judges. The service returns a decision for the agent host to enforce. That division matters: a policy decision must actually govern execution. Microsoft’s announcement describes capabilities and examples; it does not establish the effectiveness of every customer deployment.[1]
Consider an illustrative purchasing agent. It might select an approved supplier and stay within a transaction limit, yet place an order the business no longer needs. A spending rule can constrain the transaction without establishing whether the underlying commercial judgement is sound. Leaders must decide which decisions may be delegated and which require further scrutiny.
The Four-Layer Assurance model discussed in Orchestration and Command in the Age of AI provides a way to examine those responsibilities. The working AI proposes or performs the task. An independent checking AI examines its work. A runtime assurance controller has authority to pause, flag or reroute activity. Human command sets the boundaries and retains authority over consequential decisions.
These are distinct responsibilities, even where a platform combines some of their technical functions. A checking AI needs enough separation to challenge the working system, and its own limitations must be tested. Agreement between two models is not proof that a decision is correct. The controller needs enforceable authority at the boundary where an action becomes a real-world consequence.
This is an analytical comparison with Microsoft’s approach, not a claim that Microsoft has adopted the Four-Layer Assurance model. Microsoft explicitly includes human-approval requirements among the business controls its approach is intended to support.[1] The organisational question is how approval works in practice: who receives an escalation, what evidence they see and whether they can stop the action in time.
Human command also depends on competence. A person who routinely accepts an automated recommendation without understanding it provides little assurance. Organisations need named people who can question the evidence, change the permitted scope of activity and decide whether a stopped process should resume. They also need a defined response when a checker or controller becomes unavailable.
Before extending an agent’s authority, leaders should ask for evidence that consequential actions pass through the required controls and that blocked actions remain blocked. Records should connect the proposed action, the applicable policy, the decision and what actually happened. Exercises should test intervention and recovery as well as successful task completion.
Microsoft’s announcement gives leaders a timely reason to review their own arrangements. Installing runtime controls is one step; establishing who can challenge, interrupt and authorise the work is a continuing management responsibility. Orchestration and Command in the Age of AI explores that larger question: how organisations can use AI capability while keeping humans meaningfully in command.
Source
- [1] Microsoft, Command Line, 18 November 2026. Introducing agent controls: An enterprise control plane for agent behavior
More articles and information about the books: ai-humancapability.com