Making the planner aware of constraints before it plans

[SanjnaUmesh]

Hey everyone,

I have been spending some time understanding the Microsoft Agent Framework, I ended up reading quite a bit about how this evolved. I keep thinking a lot about one area - policy enforcement at the planner level.

Right now, policy/constraints enforcement (auth, data sensitivity, permissions) happens at the infrastructure layer, that is after the planner has already decided what to do. So the planner may generate a perfectly reasonable plan, a step gets blocked because it violates some constraint, and then it has to replan. Wasted tokens, extra latency, unnecessary retries.

What if the planner knew the constraints upfront?

The idea is simple, before the planner generates a plan, give it visibility into what's allowed and what's not. Things like:

• This data is sensitive, don't route it through Agent B
• This tool call needs human approval, so restructure the plan to minimize pauses
• This agent has a cost budget, pick cheaper alternatives

If the planner knows this ahead of time, it generates compliant plans from the start instead of learning through trial and error at the enforcement layer. Fewer blocked steps, fewer retries, fewer wasted LLM calls.

These are some of the patterns which I have seen come up in many papers which are trying to mitigate PIAs (eg :FIDES (https://arxiv.org/abs/2505.23643) and PRUDENTIA (https://openreview.net/pdf?id=g0aVCDY3gS)). So I was thinking the same pattern works for any constraint. Cost budgets, latency SLAs, data residency, team permissions, HITL minimization. Basically anything that can block or slow down execution if the planner sees it upfront, it plans around it.

Pretty new to this space so might be missing context on what's already been explored.