agent-developer

name	Agent Developer
description	Microsoft Agent Framework + Foundry Toolbox MCP integration specialist
tags	["python","azure","agent-framework","mcp","foundry"]

Agent Developer Skill

You are a specialist in building and configuring the AI agent backend for this project.

Tech Stack

• Python 3.11+
• Microsoft Agent Framework 1.0
• FastAPI + Uvicorn for the API gateway
• Azure AI Foundry SDK (azure-ai-projects)
• Azure Identity (DefaultAzureCredential)

Architecture

This project uses a single agent that connects to the Foundry Toolbox via one MCP endpoint. The Toolbox dispatches to individual tools (FoundryIQ, WorkIQ, custom MCP, FabricIQ) behind the scenes.

API Gateway → Single Agent → Toolbox (MCP endpoint) → Multiple Tools

Agent Modes

• Local (AGENT_MODE=local): Agent runs in-process, connects to Toolbox MCP directly via MCPStdioTool or MCPStreamableHTTPTool
• Hosted (AGENT_MODE=hosted): Agent deployed as container in Foundry Agent Service, Toolbox configured via Foundry project

Streaming Protocol

The gateway streams newline-delimited JSON events:

{"type": "message", "content": "chunk of text"}
{"type": "activity", "tool": "FoundryIQ", "status": "running", "detail": "Searching knowledge base..."}
{"type": "activity", "tool": "FoundryIQ", "status": "complete", "detail": "Found 3 results"}
{"type": "citation", "source": "doc-name", "url": "..."}

Activity events must surface individual tool names within the Toolbox, not just "Toolbox called".

File Organization

src/fibey/agent/
├── api_server.py     # FastAPI gateway
├── main.py           # Local-mode entrypoint
├── hosted.py         # Hosted-mode entrypoint
├── agent.py          # Agent definition + Toolbox MCP connection
└── prompts/
    └── system_prompt.md

Conventions

• System prompts are stored as markdown in prompts/ and loaded at runtime
• Use python-dotenv for environment configuration
• All Azure auth via DefaultAzureCredential
• Type hints on all function signatures
• Use asyncio and async/await throughout — the agent framework is async-first