Topic
AI Agents & Workflows
Give agents scoped tools, coordinate their work and reconcile effects after a lost reply or restart.

Human Approval After a Restart: One Reviewed Proposal, One Local Effect
Bind approval to the exact draft, evidence and policy, resume a pending review from an Agent Framework checkpoint in a new process, and recover a lost receipt without a second effect.

When an Agent Coordinator Has a Real Choice
Drop selector calls where the dependency graph already decides, keep a checked selector only at a real fork, and charge every attempt so a fluent loop that makes no progress still stops.

An AI Agent on SQL Server: Typed Intent, Real Read-Only, Rows Read
One question, four answers. Give a SQL agent a typed tool, a login SQL Server restricts, and measured rows read, then check every read-only claim with a command.

One Refund, Not Two: Agent Side Effects Across Timeouts and Restarts
A .NET support agent refunds through a payments API. Keys, receipts, reconciliation and a fence keep one refund per intent through lost replies, retries, restarts and reset chats.

Agent Authority You Can Prove: Reviews, Budgets and Loops That Stop
An agent that changes records and spends credits needs authority decided at the effect: a reviewer bound to the exact change, live permissions, reservations and a loop with typed endings.

Upgrade the AI Stack Without Changing What the Agent Does
Move a .NET agent from Semantic Kernel to Agent Framework in measured steps: a behaviour matrix, a library tracker, a safe cutover and Copilot Studio as a contract.

What AI Work Really Costs: Failed Attempts, Review and the Azure Bill
Count every attempt, keep missing usage unknown and divide by accepted answers. A .NET cost ledger shows when review, not tokens, decides which model is cheaper.