The article is an author-described architecture. It gives no measured performance numbers, so read it as a design pattern and not as proof of better operational outcomes.
The core idea: reasoning and execution are separate planes
The article borrows the "control plane" idea from networking. In the reasoning plane, an orchestrator and bounded domain agents correlate signals, retrieve runbooks, estimate business impact, predict failures and propose fixes. None of them holds authority to change production.
A proposal reaches the execution plane only through two things:
- A policy gate: it maps each proposed action to a risk class and an approval requirement, outside the model.
- A governed tool gateway: it is the single path to production. It enforces authorization, verifies approvals, issues scoped credentials, validates arguments and records evidence.
The practical payoff is that a hallucinated root cause becomes a bad proposal and not a bad action. A policy rule can reject it, a human can review it, or the gateway can deny it.
What makes the "one door" real
A gateway is only a boundary if agents can't go around it. The author lists several controls for that:
- No target-system credentials in the reasoning runtime.
- Deny-by-default identity policies.
- Tightly restricted network paths.
- Credentials exposed only to the tool implementation that needs them.
The article also notes that typed tools and content screening do not make prompt injection go away. An agent can still misuse an operation inside its allowed scope. The recommended answer is layers: treat ingested logs and tickets as untrusted data, use narrow tools, validate arguments, apply least privilege, require approvals, constrain the network and check outcomes after the action.
The action catalog and risk classes
Risk is decided by a versioned action catalog, not by the model's confidence. Each permitted operation carries the following fields, so the policy can be tested without an LLM:
- Tool identifier
- Target environment
- Allowed scope
- Risk class
- Approval type
- Timeout
- Preconditions and expected postconditions
- A rollback reference, where rollback is meaningful
| Risk class | Examples | Governance |
|---|---|---|
| Low | Read-only diagnostics, health checks | May auto-execute, fully logged |
| Medium | Stateless service restarts, scaling within limits | Approval per policy |
| High | Production config changes, failover, sensitive-data operations | Explicit approval, dual control where required |
The article's walkthrough shows how this plays out. A customer-payments API slows down, downstream retries climb and a queue grows. None of the signals trips an alert alone. The agents correlate them and propose diagnostics, then a restart of a stateless component. The catalog makes the restart a human-approval action, so the agent can't lower that bar by being confident.
After an approved restart, the workflow checks latency, retries, queue trend and service health. A command that returned successfully isn't treated as success. As the article puts it, an "API returned 200" is not the same as a healthy system.
Prediction gets no special authority
Prediction in this design means spotting weak-signal combinations that have historically come before failures. The article says a confidence score is not an authorization token. Predictions go through the same governed path as incidents they anticipate. It also says false positives are expected and should be measured and reduced over time.
The Microsoft mapping, and its caveats
The author says the pattern isn't tied to AWS. On Azure, he maps it like this:
- Microsoft Foundry Agent Service as the managed runtime.
- Microsoft Entra for agent identities.
- Azure AI Search for retrieval.
- Azure API Management as the tool gateway and AI gateway for model and MCP traffic.
He adds that some of these governance features are still in preview. Microsoft's documentation supports the broad mapping. Foundry lets you route Model Context Protocol (MCP) traffic through an AI gateway to control how agents access external tools. The preview has limits that matter if you plan to treat it as your production-action boundary:
- The feature is in preview, and only new MCP tools created in the Foundry portal that don't use managed OAuth are routed through the gateway.
- Existing tools are not automatically mediated by the gateway. Microsoft's troubleshooting guidance for a tool that still calls the MCP server directly is to re-create it after the gateway is connected.
- Only MCP tools are supported. SharePoint, code-first MCP tools, tools with managed OAuth and OpenAPI tools are not.
- The AI gateway doesn't log tool traces. Microsoft points admins to API Management logging for gateway traffic and to MCP server logs for tool-level detail.
Microsoft's AI gateway is also expanding. Azure API Management now has a dedicated AI Gateway tier in public preview. It was announced as free in preview, in East US 2 and Sweden Central.
The gap between the two clouds is worth noting. AWS's AgentCore Policy is built specifically for deterministic, action-level authorization at the gateway. Microsoft's current Foundry preview is stronger on routing, authentication, rate limits, IP filtering and audit logging than on per-action approval logic. If you adopt the pattern on Azure, expect to supply your own approval verification, action catalog and bypass prevention. Don't assume the preview delivers them. That is my analysis, based on the documented preview scope, not a claim made in the InfoQ article.
AWS details that carry dates
The article says AgentCore Policy has been generally available since March 2026. Its rules are written in natural language or policy-as-code and compiled to Cedar, AWS's open-source authorization language. They are attached to an AgentCore Gateway that evaluates every agent-tool request.
Since June 2026, the article says, policy can also evaluate Bedrock Guardrails. AWS's release notes describe that as a check on outputs from authorized agent actions and on inputs to gateway targets. It screens for prompt injection, harmful content and sensitive data exposure. Those checks run at the gateway layer, outside agent code. Availability is regional, so check the live region list before you plan a rollout.
AWS's own guardrails documentation adds a caution. Guardrails are non-deterministic, so the same input can score differently on separate calls. Policies themselves stay deterministic. That backs the article's advice to treat content screening as one more layer and not as a replacement for bounded authority.
Compliance: map evidence to each system
The author warns against calling a whole utility estate "CIP compliant." NERC CIP applies to the Bulk Electric System and its defined in-scope BES Cyber Systems. For those systems, the article lists:
- CIP-010-4: configuration change management and vulnerability assessments.
- CIP-004-7: personnel and access management.
- CIP-011-3: protection of BES Cyber System Information.
- CIP-008-6: incident reporting and response planning.
An agentic platform can supply evidence for applicable requirements, but it doesn't create compliance. Billing, field operations and corporate applications may fall under other regimes. NIST's AI Risk Management Framework and ISO/IEC 42001:2023 can organize the wider governance program, but they don't replace sector-specific rules.
The audit record should let you reconstruct the originating signal, retrieved evidence, recommendation, policy, human decision, executed operation, identity and validated result. The author says whether that record can be called immutable depends on the actual storage controls.
Earning autonomy in phases
The rollout runs one action class at a time:
- Read-only: observe, correlate, retrieve, predict and recommend. Move on once operators consistently accept recommendations and rejection reasons are reviewed.
- Diagnostics: move from approval-only to automatic once none ran out of scope and evidence records are complete.
- Reversible changes: introduced as approval-only until post-action validation confirms recovery and rollback has been tested.
- Selected low-impact reversible changes: automatic only within defined scope and time windows.
- High-risk changes: always explicit approval, never eligible for automatic execution.
No class advances on model confidence alone. The system owner and change-management reviewers must agree, and autonomy can be revoked as easily as it is granted.
What admins can take from it
- Inventory the doors. Find every path from an agent to a production change. If there is more than one, the gateway isn't a boundary.
- Keep the risk label out of the model. Store classification and approval rules as versioned configuration.
- Check the preview limits. If you are on Azure, confirm which tools actually route through the gateway and where traces live.
- Validate outcomes, not commands. Define the postconditions before you grant any automation.
- Treat runbooks as production data. The article describes giving each runbook a named owner and review date, and routing rejections that cite a faulty runbook back to that owner.
- Measure against your own baseline. The article tracks detection time, time to resolution, prediction precision, approval latency, policy denials and audit reconstruction time. It deliberately avoids unsupported improvement percentages.
The article is an AWS-centered reference design from one team and one environment. Its value is the pattern, which is portable: agents recommend broadly, and versioned policy plus a non-bypassable, auditable execution path decide what changes production.
References
- Governed Autonomous Intelligence for Mission-Critical Utility - infoq.com infoq.com · 2026-11-16T11:00:00+00:00
- Govern MCP Tools by Using an AI Gateway - Microsoft Foundry | Microsoft Learn learn.microsoft.com
- AI Gateway tier of API Management now in public preview techcommunity.microsoft.com