Collborative Agents with Tools & Knowledge (Graphs) using LangGraph & LangChain
- 1. Building Enterprise AI Systems with Multi-Agent Architectures, Tools & Knowledge Lucas Hänke de Cansino - AI Systems Engineer 1
- 2. 1 2 … it lacks relevant context. Building AI Systems We have all played with LLMs
- 3. Reference: https://en.wikipedia.org/wiki/Control_flow 3 1 Building AI Systems Good LLM applications follow a control flow.
- 4. 4 1 Building AI Systems This control flow forms a “chain” Reference: https://en.wikipedia.org/wiki/Control_flow
- 5. 5 1 Building AI Systems Reference: https://react-lm.github.io/ Autonomous agent-defined control flow ... using tools to execute steps.
- 6. ● Task ambiguity ● LLM non-determinism ● Tool misuse ● … and more! 6 1 Building AI Systems Poor reliability caused by Reference: https://www.reddit.com/r/LocalLLaMA/comments/1648pav/thoughts_on_autonomous_llm_agents/
- 7. 7 1 Building AI Systems Can we have both?
- 9. What is ? LangGraph applications balance agent control with agency Its core pillars support: ● Controllability: to define both explicit and implicit workflows ● Persistence: for thread-specific and cross-thread “Memory” ● Interaction: between agents, humans and external systems ● Streaming: to expose any event (or token) as it occurs LangGraph also: ● Python & Typescript implementation ● Jupyter notebooks for rapid prototyping ● Integration with Traceability and Prompt Management (LangSmith & Langfuse) 9 2 Introducing LangGraph A B C D LangGraph Studio
- 10. 10 2 Introducing LangGraph LangGraph allows for developer + LLM-defined control flows Controllability A We can have both!
- 11. 11 2 Introducing LangGraph Branches enable parallel execution of nodes to speed up overall graph operation Reference: https://langchain-ai.github.io/langgraph/how-tos/branching/ Controllability A
- 12. 12 2 Introducing LangGraph Subgraphs enable complex system design by managing states separately Reference: https://langchain-ai.github.io/langgraph/how-tos/subgraph/ Controllability A
- 13. 13 2 Introducing LangGraph Multi-agent architectures Reference: https://langchain-ai.github.io/langgraph/how-tos/persistence/ Controllability A Multi-agent network Agent Supervisor Hierarchical Agent Teams
- 14. 14 2 Introducing LangGraph ... thread-specific and cross-thread Reference: https://langchain-ai.github.io/langgraph/how-tos/persistence/ Persistent “Memory” B
- 15. 15 2 Introducing LangGraph Breakpoints enable human-in-the-loop interactions Reference: https://langchain-ai.github.io/langgraph/how-tos/human_in_the_loop/breakpoints/ Interaction C
- 16. 16 2 Introducing LangGraph First-class support for token and event-level streaming Reference: https://github.com/langchain-ai/langgraph Streaming D
- 18. LangChain’s AI abstractions and integrations make it the #1 choice for developers when building with GenAI 100K+ Apps Powered 2K+ Contributors 18 4 LangGraph within LangChain’s Ecosystem 15M+ Monthly Downloads
- 19. 19 LangGraph within LangChain’s Ecosystem LangSmith is a unified DevOps platform, purpose-built for LLM applications 300M+ Traces logged 30K+ Monthly active teams 100K+ Users signed up 4
- 20. 20 4 LangGraph within LangChain’s Ecosystem
- 21. Setting Up Useful Evaluations for Agents 21 5
- 22. 22 Feedback loops are your moat Reference: https://www.oreilly.com/radar/what-we-learned-from-a-year-of-building-with-llms-part-iii-strategy/ 5 Setting Up Useful Evaluations for Agents
- 23. 23 Setting Up Useful Evaluations for Agents Three types of feedback loops to speed up development 5
- 24. Build an agent! CoLab Run locally Notebook Use APIs Notebook Try different models Try local embeddings Try adding Self-RAG Try indexing your own docs Modify the eval set Run your own eval 24 Setting Up Useful Evaluations for Agents Choose your own adventure! 5 Try OpenCanvas (All notebooks have optional testing / eval)
- 25. 25 Get in touch! Join the LangChain Community