[Feature] First-class support for Anthropic Claude Skills in AnthropicChatCompletionClient / AssistantAgent

[xixinILYA]

Description

I'm using autogen 0.7.5 with autogen-agentchat and autogen-core and would like to use Anthropic Claude Skills together with AssistantAgent / Swarm teams.

Currently I can use Claude models via AnthropicChatCompletionClient as the model_client for an AssistantAgent, but there is no documented / first-class way to:

• pass the container.skills parameter required by Anthropic Skills, and
• set the corresponding anthropic-beta headers (e.g. code-execution-2025-08-25, skills-2025-10-02) in a convenient way from AgentChat.

Right now the only option seems to be subclassing AnthropicChatCompletionClient and overriding create(...) to inject container and headers manually.

I'd like to know what the recommended pattern is, and whether autogen could provide first-class support for Claude Skills.

---

Current behavior

• AnthropicChatCompletionClient exposes tools, tool_choice, json_output, and extra_create_args, but there is:

  • no explicit config option for skills / container, and
  • no example of using Claude Skills from AssistantAgent / Swarm.

• Using Claude as model_client for AssistantAgent works, but the agent does not automatically load or use Claude Skills that are defined in my Anthropic account.

---

Expected behavior

It would be very helpful if autogen provided one of the following:

1. A first-class config option for Claude Skills in AnthropicChatCompletionClient, e.g.:

   client = AnthropicChatCompletionClient(
       model="claude-3.7-sonnet-latest",
       api_key=...,
       skills=[
           {"type": "anthropic", "skill_id": "pptx", "version": "latest"},
           {"type": "custom", "skill_id": "skill_01AbCdEf...", "version": "latest"},
       ],
   )