Where should I integrate pre-execution verification into Foundry Agent Framework?

[foundationAIone]

Hi everyone — I’m looking for technical guidance on the correct Microsoft Foundry or Agent Framework integration point for a runtime architecture I’m testing.

The problem I’m addressing is this:

How can an agent proposal be deterministically checked, authorized, and bounded before a tool executes—and then prevented from entering persistent memory until the actual consequence is verified?

The workflow is:

proposal
→ pre-execution verification
→ authority decision
→ bounded execution
→ consequence verification
→ controlled state update

I currently have this working locally through a model-independent runtime called SE6. The LLM proposes an action, but a separate boundary returns one of:

ALLOW
REPAIR
REQUIRE_APPROVAL
PAUSE
REFUSE

Allowed actions run only through scoped adapters. The runtime then records what actually happened and permits only verified outcomes to update persistent state.

I need help understanding:

1. Which Foundry or Agent Framework extension point should intercept the proposal before tool execution?
2. What is the recommended way to preserve durable agent state across model changes?
3. Can unverified tool output be quarantined before it reaches memory?
4. Which evaluation tools should measure tokens, retries, failed executions, latency, cost, and task completion?
5. Is this lifecycle already supported as a documented Foundry pattern, or would it require custom middleware?

I have attached a short architecture diagram. I’m looking for criticism, existing references I may have missed, and preferably one engineer willing to review a minimal benchmark or prototype.

I am not assuming this is a missing Foundry capability. I am trying to determine what already exists, what can be composed from current features, and whether there is a meaningful gap worth testing.