Structured Outputs

New Features

• Removed Instructor: Agency Swarm no longer relies on the instructor. If a Pydantic model is passed, the response_format parameter will now use OpenAI's implementation. The tool creation process remains unchanged.
• strict mode: You can now enable "strict" mode when creating tools in both the BaseTool class and in your schema using ToolFactory. See Structured Outputs for more details.
• response_format: json_schema: The response format can now be set to json_schema in both the Agent class and completion methods.
• get_completion_parse: This new method allows you to pass a Pydantic model directly to the response_format method, which will then return a parsed model. If the request is refused, a RefusalError from agency_swarm.util.errors will be raised.
• Verbose mode for get_completion: By setting the verbose parameter to true, you can view intermediary messages, similar to Crew AI.
• Improved error handling for existing threads: The system will now automatically cancel any active runs within a thread and continue the conversation.

Bug Fixes

• Fixed shared state initialization when testing tools separately. #165
• Improved common error handling like connection errors and rate limits in Thread.

⚠️ Breaking Changes in BaseTool Class

• The one_call_at_a_time parameter in BaseTool is now located under ToolConfig.

• shared_state and caller_agent are now private attributes.

  from agency_swarm.tools import BaseTool
  
  class Action1(BaseTool):
      input: str = Field(...)
      # one_call_at_a_time = True  # before 
  
      class ToolConfig:
          one_call_at_a_time = True  # now
          strict = True # new
  
      def run(self):
          # self.shared_state.get('param')  # before
          self._shared_state.get('param')  # now

Full Changelog: GitHub Compare v0.2.6...v0.2.7