Agent Architectures
Architecture drives debuggability, parallelism, and cost — tangled monoliths are hard to fix.
Open studio vs office with departments vs assembly line — floor plan matters.
Visual Workflows
Start here — scroll inside each diagram frame to explore, then use + / − to zoom up to 200% if needed.
Scroll inside the frame to explore · use + / − to zoom up to 200%
ReAct Loop Detail
Scroll inside the frame to explore · use + / − to zoom up to 200%
Thought Action Observation repeated until final answer.
LangGraph State Machine
Scroll inside the frame to explore · use + / − to zoom up to 200%
Nodes edges state and checkpoints for complex branching.
Key Takeaways
- 1.ReAct: simplest loop one LLM call per step.
- 2.LangGraph: stateful graph conditional edges checkpoints.
- 3.Supervisor: router LLM delegates to specialists.
- 4.Pipeline: fixed stages retrieve plan execute verify.
Real Example
Scenario
Employee onboarding in LangGraph: nodes `verify_id` → `parse_docs` → `create_account` → `send_welcome`. Conditional edge from `parse_docs` routes to human review when OCR confidence < 0.8.
What you would do
ReAct works for a demo, but LangGraph adds checkpoints (resume after human review), conditional routing on confidence, and per-node metrics. Use `interrupt_before=['create_account']` for HITL. Monitor: node latency and retry count per node.
Practice Task
Redraw the onboarding flow as four LangGraph nodes. Add one conditional edge and one `interrupt_before` gate. Label what state each node reads and writes.
Code Walkthrough
Highlighted lines show where Agent Architectures happens in the code.
1from typing import TypedDict # import dependencies2from langgraph.graph import StateGraph, END # import dependencies3
4class OnboardingState(TypedDict): # define a data structure or component5 docs: list6 confidence: float7 account_id: str | None8
9graph = StateGraph(OnboardingState)10graph.add_node("verify_id", verify_id_fn)11graph.add_node("parse_docs", parse_docs_fn)12graph.add_node("create_account", create_account_fn)13graph.add_node("send_welcome", send_welcome_fn)14
15graph.add_conditional_edges("parse_docs", route_by_confidence) # route to different nodes based on state16graph.add_edge("verify_id", "parse_docs")17graph.add_edge("create_account", "send_welcome")18app = graph.compile(interrupt_before=["create_account"])Cheat Sheet
Quick recap
quick ref- •ReAct = simple interpretable loop
- •LangGraph = explicit state machine
- •Supervisor = delegation pattern
- •Pipeline = fixed stage order
- •Pick based on task structure
Common Mistakes
- ✕Skipping evaluation for Agent Architectures before production
- ✕No logging or tracing around agent architectures steps
- ✕Ignoring cost and latency implications