Foundry Agent Development

Foundry Agent Development Skill

This skill provides guidance for developing AI agents using Azure AI Foundry within the Foundry Agent Accelerator project.

Project Context

This is a FastAPI + React application that creates persistent AI agents in Azure AI Foundry. Agents are:

• Visible in the Azure AI Foundry portal
• Version-controlled (new versions created on config changes)
• Configurable via local files OR the portal

Key Files

• System Prompt - Agent personality and instructions
• Agent Configuration - Tools and capabilities
• Main Application - Agent initialization
• API Routes - Chat endpoint

Configuration Modes

LOCAL Mode (default)

Agent configured via local files:

AGENT_CONFIG_SOURCE=local

• Edit src/api/prompts/system.txt for personality
• Edit src/agent.yaml for tools
• Restart creates new version if config changed

PORTAL Mode

Agent configured in Azure portal:

AGENT_CONFIG_SOURCE=portal

• Local files ignored
• Manage agent entirely in Foundry portal

Available Agent Tools

Tool Configuration Pattern

# src/agent.yaml
tools:
  code_interpreter:
    enabled: true
  
  bing_search:
    enabled: true
    connection_name: "my-bing-connection"  # Must match Foundry

System Prompt Best Practices

1. Define identity clearly: "You are [name], a [role] for [purpose]"
2. Specify personality traits: Friendly, professional, technical, etc.
3. List capabilities: What the agent CAN do
4. Set boundaries: What the agent should NOT do
5. Include response format guidelines

Azure AI SDK Patterns

from azure.ai.projects import AIProjectClient
from azure.ai.projects.models import (
    PromptAgentDefinition,
    CodeInterpreterTool,
    BingGroundingAgentTool,
)
from azure.identity import DefaultAzureCredential

# Initialize client
credential = DefaultAzureCredential()
project_client = AIProjectClient.from_connection_string(
    conn_str=os.getenv("AZURE_EXISTING_AIPROJECT_ENDPOINT"),
    credential=credential
)

# Create/update agent with versioning
agent = project_client.agents.create_version(
    definition=PromptAgentDefinition(
        name=agent_name,
        model=model_name,
        instructions=system_prompt,
        tools=tools_list
    )
)

Streaming Response Pattern

The chat endpoint uses Server-Sent Events (SSE):

async def generate_response():
    async for chunk in agent_response:
        yield f"data: {json.dumps({'type': 'message', 'content': chunk})}\n\n"
    yield f"data: {json.dumps({'type': 'stream_end'})}\n\n"

return StreamingResponse(generate_response(), media_type="text/event-stream")

Version Management

Changes to config create new agent versions:

# Hash detection prevents version spam
config_hash = hashlib.md5(json.dumps({
    "instructions": system_prompt,
    "tools": tools_config,
    "model": model_name
}).encode()).hexdigest()

# Only create_version if hash changed
if config_hash != previous_deployment_hash:
    agent = project_client.agents.create_version(...)

Troubleshooting

• Agent not creating: Check AZURE_AI_AGENT_NAME and endpoint
• Tools not working: Verify connection names in Foundry portal
• Version spam: Check for hidden whitespace changes in config files
• Auth errors: Run az login and verify subscription access