LangGraph - Multi-Agent Workflows
AI Summary
Summary: Lang Graph Library Introduction
- Introduction to Lang Graph:
- Presented by Harrison from Lang chain.
- Lang graph is a new library for creating multi-agent workflows as graphs.
- Allows dynamic creation of agent-like workflows, running language models in loops with various structures.
- Comparable to state machines with nodes (agents) and edges (communication between agents).
- Using L Graph:
- Define graphs to track state over time.
- Add nodes and edges to represent agents and their interactions.
- Can be used like other Lang chain tools.
- Multi-Agent Workflow Variants:
- Multi-Agent Collaboration:
- Agents share a state of messages, adding to a global message state over time.
- Example includes a researcher agent and a chart generator agent.
- Workflow involves researcher querying, calling tools, and chart generator visualizing results.
- Agent Supervisor:
- Supervisor routes tasks between independent agents.
- Agents work independently and return final answers to the supervisor.
- Supervisor only sees final outputs, not the internal process of agents.
- Hierarchical Agent Team:
- Similar to Agent Supervisor but with sub-teams having their own supervisors.
- Each agent node is a supervisor-agent setup, creating a more complex hierarchy.
- Practical Implementation:
- Demonstrated using Lang Smith for debugging.
- Showed how to define helper functions, create agents, and set up graphs.
- Walked through examples of each workflow variant, including the setup and execution of tasks.
- Highlighted the benefits of compartmentalizing functions within a system.
- Conclusion:
- Multi-agent workflows help in organizing complex tasks and provide a clear mental model for system functions.
- L graph offers flexibility in creating various multi-agent workflows.
- Encourages exploration of other potential multi-agent workflow configurations with L graph.