Add a hosted-agent sample: Anthropic (Claude) via APIM with Agent Framework

Problem

samples/python/hosted-agents/agent-framework/responses/01-basic is the only Agent Framework "connect to a model" sample, and it uses FoundryChatClient, which speaks the OpenAI Responses protocol to a Foundry-deployed model. Claude models exposed through Azure API Management (or Azure AI Foundry directly) are Anthropic Messages-protocol only — there is no OpenAI-compatible route — so FoundryChatClient can never reach them. There's currently no sample in this repo showing how to wire a hosted Agent Framework agent to an Anthropic model.

Proposal

Add a new sample, e.g. samples/python/hosted-agents/agent-framework/responses/14-anthropic-apim, using AnthropicFoundryClient from the agent-framework-anthropic package (built on the anthropic SDK's AsyncAnthropicFoundry) instead of FoundryChatClient:

from agent_framework import Agent
from agent_framework_anthropic import AnthropicFoundryClient
from agent_framework_foundry_hosting import ResponsesHostServer

client = AnthropicFoundryClient(
    model=os.environ["ANTHROPIC_MODEL_NAME"],       # e.g. claude-sonnet-5
    base_url=os.environ["APIM_ANTHROPIC_BASE_URL"],  # https://<apim>.azure-api.net/anthropic
    api_key=os.environ["APIM_SUBSCRIPTION_KEY"],
)
agent = Agent(client=client, instructions="...")
ResponsesHostServer(agent).run()

api_key is sent as the standard Anthropic SDK x-api-key header, matching what an APIM Anthropic-passthrough policy typically expects. azure.yaml needs no ai-project.deployments block (no Foundry-deployed model is used) — the ai-project service stays for hosting/identity, and ANTHROPIC_MODEL_NAME/APIM_ANTHROPIC_BASE_URL/APIM_SUBSCRIPTION_KEY are declared as agent env vars instead.

Foundry's hosted-agent session/conversation/history management is unaffected by the chat-client swap — ResponsesHostServer owns that, not the chat client — so multi-turn conversations work exactly as in 01-basic.

Try it now

I published a standalone, runnable version of this sample (same file layout as 01-basic) at https://github.com/lordlinus/foundry-agent-anthropic-apim — it works with the exact same azd ai agent init -m <url> flow used for the official samples:

mkdir my-anthropic-apim-agent && cd my-anthropic-apim-agent
azd ai agent init -m https://github.com/lordlinus/foundry-agent-anthropic-apim/blob/main/azure.yaml

Verified end-to-end: azd ai agent run/invoke --local and a full azd provision/azd deploy/azd ai agent invoke deployment against a live Foundry project, including multi-turn session continuity.

About a pull request

I have the full sample ready to submit (matching 01-basic's file layout: azure.yaml, README.md, AGENTS.md, CLAUDE.md, src/agent-framework-agent-anthropic-apim/{Dockerfile,main.py,requirements.txt,.env.example,.azdignore,.dockerignore}, identical to the repo linked above), but I could not open a working pull request:

• I don't have branch-creation permission on microsoft-foundry/foundry-samples, so per CONTRIBUTING.md a same-repo-branch PR isn't possible for me.
• My existing fork (lordlinus/foundry-samples) predates a history change upstream, so GitHub reports "no history in common" and refuses to open a PR from it at all — re-forking would be needed first.

I'm happy to push the sample to a same-repo branch if a maintainer can grant branch-creation access, or to send the file contents directly if that's easier to land.

Compare samples/python/hosted-agents/agent-framework/responses/01-basic with the proposed azure.yaml and src/agent-framework-agent-anthropic-apim files, especially main.py and requirements.txt. Start by checking the AnthropicFoundryClient and ResponsesHostServer entry points, then run azd ai agent run and invoke --local. Done means the sample provisions and deploys with the three documented environment variables and preserves multi-turn hosted-agent sessions.