Anthropic integration for Microsoft Agent Framework

Warnings

• breaking The `agent-framework-anthropic` package is currently in a pre-release (beta) state. Breaking changes may occur frequently, especially as the broader Microsoft Agent Framework evolves towards stable releases.
  Fix: Refer to the official `microsoft/agent-framework` GitHub repository and documentation for specific migration guides and changelogs before upgrading. Test thoroughly.
• gotcha The core `agent-framework` package often requires the `--pre` flag during `pip install` to get the latest compatible pre-release versions, as it is also under rapid development. Forgetting this can lead to version conflicts or outdated components.
  Fix: Always install `agent-framework` with `pip install agent-framework --pre` to ensure compatibility with the latest `agent-framework-anthropic` releases.
• gotcha API keys (e.g., `ANTHROPIC_API_KEY`) should never be hardcoded or committed to source control. They must be managed securely, preferably through environment variables or a dedicated secret management service.
  Fix: Set `ANTHROPIC_API_KEY` as an environment variable (e.g., `export ANTHROPIC_API_KEY='sk-ant-...'`). The `AnthropicClient` will automatically detect it.
• breaking A critical vulnerability has been identified in Anthropic's Model Context Protocol (MCP), which could allow arbitrary command execution on servers running MCP. Since Microsoft Agent Framework integrates with MCP, this is a significant security concern.
  Fix: Monitor official announcements from Anthropic and Microsoft regarding patches or mitigation strategies for the MCP vulnerability. Update dependencies as soon as fixes are available.
• gotcha Long-running AI agents built with Claude models can face challenges with memory and context windows, potentially leading to agents 'forgetting' instructions or previous conversational turns over extended interactions.
  Fix: Implement explicit memory management strategies or leverage features like 'initializer agents' and 'coding agents' for incremental progress, as discussed by Anthropic for their SDK.
• gotcha Integrating deeply with Anthropic's infrastructure or the Microsoft Agent Framework may lead to increased vendor lock-in, making it harder to switch underlying LLM providers or agent frameworks in the future.
  Fix: Evaluate the trade-offs between ease of integration and vendor flexibility. Design your agent architecture with modularity in mind where possible to abstract away provider-specific implementations.
• gotcha Costs for Anthropic models vary significantly by model (e.g., Claude Haiku vs. Opus) and token usage (input/output). Inefficient prompting or long contexts can lead to unexpectedly high costs.
  Fix: Follow Anthropic's best practices for prompt engineering, choose appropriate models for the task, monitor token usage, and optimize prompts for conciseness and clarity.

Install

• pip install agent-framework --pre Install core Agent Framework (pre-release)
• pip install agent-framework-anthropic Install Anthropic integration

Imports

• AnthropicClient

  from agent_framework.anthropic import AnthropicClient

Quickstart

This quickstart demonstrates how to initialize an `AnthropicClient` and create a basic AI agent. It assumes the `ANTHROPIC_API_KEY` environment variable is set for authentication with the Anthropic API. The agent is then used to answer a simple question.

import asyncio
import os
from agent_framework.anthropic import AnthropicClient

async def main():
    # Ensure ANTHROPIC_API_KEY is set in your environment variables.
    # For a quick test, you can uncomment and set it directly, but avoid in production.
    # os.environ['ANTHROPIC_API_KEY'] = 'your_anthropic_api_key_here'

    if not os.environ.get('ANTHROPIC_API_KEY'):
        print("Error: ANTHROPIC_API_KEY environment variable is not set.")
        print("Please set it before running the quickstart.")
        return

    client = AnthropicClient()

    # Create an agent using the Anthropic client
    agent = client.as_agent(
        name="HelpfulAssistant",
        instructions="You are a helpful assistant that answers questions accurately.",
    )

    # Run the agent with a prompt
    response = await agent.run("What is the capital of France?")
    print(f"Assistant: {response}")

if __name__ == "__main__":
    asyncio.run(main())