Microsoft Agent Framework
When to use this skill
- When you need to design an agent system maintained at the organization/team level
- When you need separation of responsibilities, governance, and auditable execution logs
- When you need a structure that is operable and scalable, not a one-off demo
Instructions
Step 1) Finalize the role separation model
- Separate the roles of Planner / Executor / Reviewer / Approver
- Define the input/output contracts (schemas) between roles first
- Select the work groups that require an approval stage (Approver)
Step 2) Define the control flow
- Distinguish synchronous/asynchronous paths
- Set up compensation transactions or rollback strategies on failure
- Codify SLAs (response time/retries/abort conditions)
Step 3) Set policy boundaries
- Restrict data access scope (read/write/external transmission) per role
- Apply sensitive information masking/log retention policies
- Separate the scope of automatic execution from the scope of human approval
Step 4) Build operational observability
- Assign a trace id to each step
- Classify failure causes (policy, tool, model, input)
- Maintain scenarios for periodic regression checks
Examples
- Request: "I want to build an agent operating system that includes a cross-department approval flow"
- Applied: Configured a Planner→Executor→Reviewer→Approver pipeline
- Result: Secured an auditable approval history and failure recovery paths
Best practices
- Role separation should center on permission boundaries, not on prompt style.
- Limit automation without approval to low-risk tasks.
- Prohibit optimization without observability, and build the logging system first.
References
