LangGraph (your stack)

Mental model: stateful graph with nodes and edges.

• Nodes are functions (often wrapping an LLM call or tool call) that read and write the shared state.
• Edges are transitions — either fixed (A → B) or conditional (A → router → B or C or end).
• Cycles allowed — that's the whole point vs. a DAG framework. Agent loops are literally cycles.
• Checkpointer persists state for resume / HITL / time-travel.

When to reach for LangGraph:

• You need branching, loops, conditional routing
• You need to inspect/persist/replay state
• Multi-agent with explicit coordination

When LangGraph is overkill: a linear chain (retrieve → generate), simple ReAct loop — just use a plain function or LangChain.

Google ADK (Agent Development Kit)

Google's official framework for agents on Vertex AI. Worth name-dropping in the interview since you're going for Google Cloud.

• Multi-agent native, hierarchical delegation as a first-class pattern.
• Built-in evals, tracing (Cloud Trace integration), deployment to Vertex AI Agent Engine.
• Comes with sub-agent, tool, and workflow agent abstractions.

Canonical phrasing: "ADK is to Vertex AI Agent Engine what LangGraph is to a generic Python runtime — the framework is tighter coupled to the deployment target, which buys you tracing, evals, and managed deployment out of the box."

CrewAI

Role-based abstraction (Agent with role/goal/backstory, Task, Crew). Higher level than LangGraph — opinionated patterns, less control. Good for fast prototyping; less common in serious production deployments.

ADK vs LangGraph — the comparison they'll ask