python-v0.6.4
Breaking Changes📦 autogenView on GitHub →
⚠ 2 breaking✨ 5 features🐛 6 fixes🔧 10 symbols
Summary
This release improves GraphFlow state persistence and termination logic, adds tool override capabilities to Workbench implementations, and expands model support for Claude and Qwen2.5-VL.
⚠️ Breaking Changes
- The inner StopAgent has been removed from GraphFlow. The last message will no longer come from StopAgent; instead, check the stop_reason field in TaskResult.
- GraphFlow now retains execution state after termination conditions are hit, requiring an explicit finish (no available next agents) to reset state automatically.
Migration Steps
- Update code that relies on receiving a final message from StopAgent to instead inspect the 'stop_reason' field in the 'TaskResult' object.
- Review GraphFlow implementations to ensure logic accounts for state retention between tasks until the graph fully completes.
✨ New Features
- McpWorkbench and StaticWorkbench now support overriding tool names and descriptions for client-side optimization.
- Added reflection support for Claude models in AssistantAgent.
- Added support for Qwen2.5-VL model.
- Added DuckDuckGo Search Tool and Agent in AutoGen Extensions (Note: subsequent PR in same release notes indicates removal/reversion).
- Added script for automatic API documentation generation.
🐛 Bug Fixes
- Fixed GraphFlow to support multiple task executions without requiring an explicit reset.
- Fixed GraphFlowManager termination to prevent _StopAgent from polluting the conversation context.
- Fixed function calling support for Llama 3.3.
- Updated GitHub Models URL to the new endpoint.
- Modified SingleThreadedAgentRuntime to use subclass checks for factory_wrapper instead of equality.
- Removed otel-semconv package from core dependencies to streamline requirements.
🔧 Affected Symbols
GraphFlowGraphFlowManagerStopAgent_StopAgentTaskResultMcpWorkbenchStaticWorkbenchAssistantAgentSingleThreadedAgentRuntimeDuckDuckGoSearchTool