Create Agent with Microsoft Agent Framework

Build AI agents, agentic apps, and multi-agent workflows using Microsoft Agent Framework SDK.

Quick Reference

When to Use This Skill

Use when the user wants to:

• Create a new AI agent or agentic application
• Scaffold an agent with tools (MCP, function calling)
• Build multi-agent workflows with orchestration patterns
• Add HTTP server mode to an existing agent
• Configure F5/debug support for VSCode

Defaults

• Language: Python
• SDK: Microsoft Agent Framework (pin version 1.0.0b260107)
• Server: HTTP via Azure AI Agent Server SDK
• Environment: Virtual environment (create or detect existing)

References

💡 Tip: For advanced patterns (Reflection, Switch-Case, Fan-out/Fan-in, Loop, Human-in-Loop), search microsoft/agent-framework on GitHub.

MCP Tools

This skill delegates to microsoft-foundry MCP tools for model and project operations:

Creation Workflow

1. Gather context (read agent-as-server.md + debug-setup.md + code samples)
2. Select model & configure environment
3. Implement agent/workflow code + HTTP server mode + .vscode/ configs
4. Install dependencies (venv + requirements.txt)
5. Verify startup (Run-Fix loop)
6. Documentation

Step 1: Gather Context

Read reference files based on user's request:

Always read these references:

• Server pattern: agent-as-server.md (required — HTTP server is the default)
• Debug setup: debug-setup.md (required — always generate .vscode/ configs)

Read the relevant code sample:

• Code samples: agent-samples.md, workflow-basics.md, workflow-agents.md, or workflow-foundry.md

Model Selection: Use microsoft-foundry skill's model catalog to help user select and deploy a model.

Recommended: Search microsoft/agent-framework on GitHub for advanced patterns.

Step 2: Select Model & Configure Environment

Decide on the model BEFORE coding.

If user hasn't specified a model, use microsoft-foundry skill to list deployed models or help deploy one.

ALWAYS create/update .env file:

FOUNDRY_PROJECT_ENDPOINT=<project-endpoint>
FOUNDRY_MODEL_DEPLOYMENT_NAME=<model-deployment-name>

• Standard flow: Populate with real values from user's Foundry project
• Deferred Config: Use placeholders, remind user to update before running

Step 3: Implement Code

All three are required by default:

1. Agent/Workflow code: Use gathered context to structure the agent or workflow
2. HTTP Server mode: Wrap with Agent-as-Server pattern from agent-as-server.md — this is the default entry point
3. Debug configs: Generate .vscode/launch.json and .vscode/tasks.json using templates from debug-setup.md

⚠️ Warning: Only skip server mode or debug configs if the user explicitly requests a "minimal" or "no server" setup.

Step 4: Install Dependencies

1. Generate/update requirements.txt

# pin version to avoid breaking changes

# agent framework
agent-framework-azure-ai==1.0.0b260107
agent-framework-core==1.0.0b260107

# agent server (for HTTP server mode)
azure-ai-agentserver-core==1.0.0b10
azure-ai-agentserver-agentframework==1.0.0b10

# debugging support
debugpy
agent-dev-cli

2. Use a virtual environment to avoid polluting the global Python installation

⚠️ Warning: Never use bare python or pip — always use the venv-activated versions or full paths (e.g., .venv/bin/pip).

Step 5: Verify Startup (Run-Fix Loop)

Enter a run-fix loop until no startup errors:

1. Run the main entrypoint using the venv's Python (e.g., .venv/Scripts/python main.py on Windows, .venv/bin/python main.py on macOS/Linux)
2. If startup fails: Fix error → Rerun
3. If startup succeeds: Stop server immediately

Guardrails:

• ✅ Perform real run to catch startup errors
• ✅ Cleanup after verification (stop HTTP server)
• ✅ Ignore environment/auth/connection/timeout errors
• ❌ Don't wait for user input
• ❌ Don't create separate test scripts
• ❌ Don't mock configuration

Step 6: Documentation

Create/update README.md with setup instructions and usage examples.

Error Handling