Microsoft Agent Framework: agents and workflows

Overview

• Microsoft Agent Framework, or MAF, provides agent abstractions and explicit workflows.
• An agent handles model-driven interaction and tool use; a workflow connects agents and ordinary functions through controlled paths.
• Microsoft describes it as the successor bringing together ideas from AutoGen and Semantic Kernel.

Why this matters

Enterprise applications often need both adaptable language behavior and explicit coordination integrated with their platform.

An everyday example

A company employs specialists but still uses an official case-handling procedure.

A technical example

A claim-intake agent extracts missing information; workflow executors validate fields, wait for review, and invoke the claim service.

How it works

1. Define agents/tools
2. Define typed workflow inputs and steps
3. Connect execution paths
4. Configure state, review, and hosting
5. Run and monitor.

Architecture and flow

See the conceptual workflow below.

The insurance assistant example

An Azure/.NET-centered insurer uses MAF for agent interactions and explicit claim-preparation coordination.

When to use it

When Microsoft ecosystem integration, supported language tooling, and enterprise workflow needs fit your team.

When to avoid it

Do not replace a working TypeScript graph solely for Microsoft branding; validate language support, feature maturity, and maintenance impact first.

Interview answer

“MAF separates model-driven agents from explicit workflows. I use agents for language ambiguity and workflow steps for controlled validation, coordination, and review.”

Benefits, trade-offs, and production considerations

• Like LangGraph, explicit orchestration makes transitions reviewable.
• Verify exact persistence, hosting, replay, and checkpoint guarantees before relying on them.
• The framework does not replace domain authorization or idempotency.
• Microsoft Agent Framework overview.

flowchart LR
  I[Authenticated input] --> A[Intake agent: interpret language]
  A --> V[Function: validate fields and rules]
  V --> H[Workflow: wait for approved draft]
  H --> C[Function: idempotent claim service]
  C --> E[Agent or template: explain verified result]

• Input contains verified identity.
• The agent interprets wording.
• Validation checks business facts.
• The workflow controls when execution can continue.
• The claim service commits an authorized transaction.
• The final step explains the actual result.

Conceptual MAF design, not version-specific SDK syntax: define an intake agent returning ClaimDraft, a validator returning ValidationResult, a review step returning an authenticated ApprovalRecord, and a commit executor returning ClaimReceipt. The durable workflow carries IDs and versions between them. A claim draft is only a proposal; approval belongs to a trusted record and commit belongs to the claim backend.