Langchain, langgraph, langgraph-GUI, pydanticAI, CrewAI and many other frameworks are becoming increasing complex for me to learn the framework only to be updated with each release. And the prebuild chain and state management are pretty small functionality which with fundamental understanding can written without much dependency. Combining my understanding to focus on the 20% code to get 80% of the functionality without replying any specific dependencies(inference and pydantic are must, rest are based on agents).
- ✅ Prompt template (components, context ordering and tags for clear segregation)
- ⬜ System
-
- ✅ Persona and Instructions
-
- ⬜ Toolkits and instructions
-
- ⬜ Multimodal input
- ⬜ Agent structure
-
- ✅ ReAct structure for now. Need to find a way to use and implement different structures(Self-Reflection, CoT).
-
- ⬜ Need to find better communication protocol between agents (MCP vs google A2A)
-
- ⬜ Multimodal processing
-
- ⬜ Structured response for more complex workflows to send additional state elements
- ⬜ State implementation (Currently this is only to provide the structure for implementation. No state. Or state management)
- ⬜ Async implementation (I don't understand this completely yet)
- ⬜ Different tools/toolkits integration (Leverage CoT with MCP)
- ⬜ PyAutoGUI, playwright, for developer flow replication