# Host test plan for a hosted policy decision

**Version:** 1.0
**Last updated:** 2026-11-20
**Source:** KLA Digital (kla.digital/blog/microsoft-agent-controls-host-enforcement)

---

## About this plan

A hosted policy service decides; the agent host enforces. This plan tests the host side of that
split for any hosted decision service, including Microsoft agent controls and KLA. Run it before
production and after any change to the host, the policy or the service version.

Written from the sources below, read on 20 November 2026:

- Microsoft, *Introducing agent controls: An enterprise control plane for agent behavior*
  (18 November 2026).
  https://commandline.microsoft.com/introducing-agent-controls-an-enterprise-control-plane-for-agent-behavior/
- Microsoft Learn, *Agent controls API and operations* (preview).
  https://learn.microsoft.com/en-us/azure/foundry/agent-controls/api-and-operations
- Microsoft Learn, *What are agent controls?* (preview).
  https://learn.microsoft.com/en-us/azure/foundry/agent-controls/overview
- Microsoft Agent Framework, *Agent Hooks*.
  https://learn.microsoft.com/en-us/agent-framework/agents/agent-hooks

## The tests

| # | Case | What to do | Pass when |
| --- | --- | --- | --- |
| 1 | Allow | Send a call the policy allows | Dispatched with exactly the evaluated values |
| 2 | Deny | Send a call the policy denies | Not dispatched; denial recorded with policy version |
| 3 | Dispatch matches evaluation | Change one argument after the verdict, before dispatch | Changed call re-evaluated or refused |
| 4 | Review requested | Policy returns a deny carrying an approval object | Action stays blocked until your workflow records a decision bound to the evaluated values |
| 5 | Review rejected or expired | Reject, or let the review expire | Action never dispatched; outcome recorded |
| 6 | Unconditional deny beside approval | One policy asks for review, another denies | Action stays blocked; the approval cannot lift the deny |
| 7 | Service unavailable | Force HTTP 503 and a client timeout | Action stays blocked; retries stop at a deadline; event recorded |
| 8 | Rate limited | Force HTTP 429 with Retry-After | Action stays blocked; Retry-After honoured |
| 9 | Malformed verdict | Return an inconsistent or unparseable decision | Action stays blocked |
| 10 | HTTP 200 with deny | Return 200 and a deny | Action stays blocked (a 200 means the evaluation completed) |
| 11 | Transform | Return a transformed tool argument | Replacement validated against your schema, applied once, dispatched value recorded |
| 12 | Aggregate limit | Two transfers under the per-call limit that together exceed a daily limit | Second denied because the request context carried the authoritative running total |
| 13 | Uncovered path | Run a tool the model provider executes server-side, or a code path without the hook | Listed as uncovered in your control inventory, with its own control |
| 14 | Correlation | Pick one dispatched action from last week | You can show the policy version, the event sent, the arguments dispatched, the observed outcome and who approved it |

## What to record per evaluation

- Event identifier, sequence and timestamp sent to the service
- Policy identifier and version returned
- Decision, reason, warnings and any approval or transform
- Arguments actually dispatched, or the refusal
- Approval decision record, if any, bound to the evaluated values
- Observed outcome in the destination system
- Protection of the record (signature, append-only store), since the service's published receipt
  material shows no signature or verification procedure

## Appendix: checks on 20 November 2026

Run with curl from a Linux host. We had no preview account and ran no evaluation.

| Check | Result |
| --- | --- |
| Article-linked samples, `github.com/microsoft-foundry/foundry-samples/tree/main/samples/cli/agent-controls` | HTTP 404; the GitHub API also returned 404 for the repository |
| Announcement, Learn overview, API and operations, quickstart, policy authoring, Agent Framework Agent Hooks, `responsibleai/agent-control-spec`, `responsibleai/agent-hooks` | HTTP 200 |
| Policy selector | Announcement: full ARM resource ID as URL-encoded `policyId` query parameter. Learn API: "Put policyId in the JSON body, not the query string. Use the published policy's name". |
| API versions | Unified Moderate `2026-09-01-preview` in all pages; ARM `2026-11-15-preview` in quickstart and policy authoring; manifest `0.4.0-alpha.1` |
| Event sources | Unified Moderate `2026-09-01-preview`: `input`, `output`, `pre_tool_call`, `post_tool_call`. Agent Hooks defines eight interception points. |
| Approval | "A review decision is a deny until your host resolves it." "The approval field does not open a review queue or grant permission." "Bind any approval to the exact action and evaluated values." |
| Errors | 429: honour Retry-After, keep the action blocked. 500, 503, transport error or timeout: keep the action blocked, retry within a bounded deadline. |
| Receipt | Announcement: each evaluation returns a receipt identifying the policies and versions applied. Learn response fields: `verdict`, content, `acsVerdict` (decision, reason, message, warnings, approval, transform, evidence, harmResults). No receipt field, signature or verification procedure shown. |
| Transfer example | Printed Rego defaults to deny (`tool_call_not_permitted`) and allows `transfer_funds` only for USD, 0 < amount ≤ 10000. An over-limit transfer is denied with no approval object. |
| Framework support | Agent Framework: Agent Hooks experimental in Python, not yet in .NET or Go; provider-hosted tools cannot be blocked at `pre_tool_call`. |
