Selective Recovery for Microsoft Agent Framework Workflows

[Oluseyi Christopher Olusola (Tarchy01)]

Hi Agent Framework team,

I’m Oluseyi, founder of Consistency.

We’ve built a recovery layer for AI agents and multi-step workflows that focuses on a specific problem after partial failure: deciding which completed work is still valid and which downstream work actually needs to run again.

We have already tested Consistency on Microsoft Agent Framework DAG workflows, including multi-fault recovery scenarios.

The basic idea is:

When part of a workflow fails, Consistency examines the surviving state, preserves the work that is still valid, repairs the failed boundary, invalidates only the downstream state affected by that failure, and reruns only what needs reconstruction.

We also tested the same recovery approach in a live OpenHands agent workflow:

Metric	Restart	Consistency
Model calls	3	2
Prompt tokens	1,961	223
Total tokens	2,043	236
External tool calls	3	1
Recovery time	11.261s	6.451s
Final result	HydroLink	HydroLink

Same final answer, with less repeated work.

Demo:
https://consistency-runtime.netlify.app

Given Agent Framework’s work on durable, restartable workflows, I think Consistency could add another layer of recovery by helping determine which completed steps are still safe to keep after a failure and which ones need to run again.

I’d be interested in feedback from the Agent Framework team on whether this could fit naturally into the framework, and whether there might be room for a small integration or design-partner experiment.

Best,
Oluseyi Christopher Olusola
Founder, Consistency