Build a Bounded Microsoft Agent Framework Workflow
Build a Microsoft Agent Framework workflow with typed steps, narrow tools, review boundaries, and tested checkpoints.
Choose the supported Python or .NET package version and distinguish Microsoft Agent Framework from the separate Microsoft 365 Agents Toolkit. Start with a workflow that takes a synthetic request, proposes a read-only lookup or reversible action, and returns a typed result. Represent state explicitly. Include input identity, extraction status, proposed action, approval reference where needed, and any operation ID used for recovery. ## Key Points - [Microsoft Agent Framework workflows](https://learn.microsoft.com/en-us/agent-framework/journey/workflows) - [Workflow checkpoints](https://learn.microsoft.com/en-us/agent-framework/user-guide/workflows/checkpoints?pivots=programming-language-python) - [Human-in-the-loop workflows](https://learn.microsoft.com/en-us/agent-framework/workflows/human-in-the-loop)
skilldb get ai-agent-orchestration-skills/microsoft-agent-frameworkFull skill: 31 linesBuild a Bounded Microsoft Agent Framework Workflow
Define state and control flow before adding a model
Choose the supported Python or .NET package version and distinguish Microsoft Agent Framework from the separate Microsoft 365 Agents Toolkit. Start with a workflow that takes a synthetic request, proposes a read-only lookup or reversible action, and returns a typed result. Represent state explicitly. Include input identity, extraction status, proposed action, approval reference where needed, and any operation ID used for recovery.
Choose a deterministic path for validation and routing. Use an agent step only where interpretation is valuable, such as mapping text into a bounded request schema. Keep tool arguments explicit and check them before invocation. Limit each tool to the required capability and validate its result before updating typed workflow state. A model's confidence or fluent explanation does not substitute for authorization or source evidence.
Add review and resume boundaries
Use the framework's documented human review mechanism when an action requires a decision. Present the exact action, relevant input version, permissions, and expected effect to the reviewer. Record the resulting decision separately from a model-generated recommendation. A review for an earlier input cannot authorize a changed payload.
Use a resume checkpoint only where stored state has enough information to continue safely. Give any external action a stable operation identifier and reconcile state before replaying uncertain work. Separate provider failures, denied decisions, validation errors, and completed steps. If a checkpoint can contain user content or tool output, secure it under the project's retention policy. Log correlation and step outcomes without exposing credentials or sensitive text.
Test the workflow as a state machine
Use fake model and tool adapters for deterministic state tests, then exercise one approved integration fixture against the pinned SDK. Cover normal completion, invalid structured output, tool timeout, rejected review, stale review, and checkpoint recovery. For a repeated resume, confirm that a previous successful action is not repeated. Verify streamed progress represents intermediate state rather than a completed operation.
Save the package versions, workflow diagram or route list, test traces, and step outcomes. Distinguish the unit scenario from a provider run with the real framework and tool adapter. Keep production credentials and connector changes outside the fixture.
References
Install this skill directly: skilldb add ai-agent-orchestration-skills
Related Skills
multi-agent-systems
Orchestrating multiple AI agents working together: supervisor patterns, swarm architecture, handoff protocols, agent-to-agent communication, and agent specialization. Covers practical patterns for splitting complex tasks across coordinated agents, managing shared state, and routing work to the right specialist agent.
tool-calling
Implementing tool and function calling across Claude, OpenAI, and Gemini APIs. Covers schema design best practices, parallel tool calls, error handling, tool result formatting, dynamic tool registration, and patterns for building composable tool sets that agents can use reliably.
agent-architecture
Core patterns for building AI agent systems: the observe-think-act loop, ReAct pattern implementation, tool-use cycles, memory systems (short-term and long-term), and planning strategies. Covers how to structure an agent's main loop, manage state between iterations, and wire together perception, reasoning, and action into a reliable autonomous system.
agent-error-recovery
Handling failures in AI agent systems: retry strategies with backoff, fallback tools, graceful degradation, human-in-the-loop escalation, stuck-loop detection, and context recovery after crashes. Covers practical patterns for making agents robust against tool failures, API errors, and reasoning dead-ends.
agent-evaluation
Testing and evaluating AI agents: trajectory evaluation, task completion metrics, tool-use accuracy measurement, regression testing, benchmark suites, and A/B testing agent configurations. Covers practical approaches to measuring whether agents are working correctly and improving over time.
agent-frameworks
Comparison of major AI agent frameworks: LangGraph, CrewAI, AutoGen, Semantic Kernel, and Claude Agent SDK. Covers when to use each framework, their trade-offs, core patterns, practical setup examples, and migration strategies between frameworks.