Agent Architecture Playground

Six LLM agent architectures on one shared runtime — a visual playground that shows how each one plans, uses tools, and self-corrects, not just the final answer.

Agents

Architecture	Flow	Stack
Single LLM	input → structured output	Pydantic
Tool Calling	decide → call tool → answer	LangGraph
ReAct	act → observe → repeat	LangGraph · Pinecone · Tavily
Plan-and-Execute	plan → execute → validate → replan	Pydantic · LangChain
Reflection	draft → critique → revise	LangGraph
Reflexion	answer → reflect → search → revise	Tavily

Python · Streamlit · LangGraph / LangChain · OpenAI · LangSmith.

Demos

Single LLM — review → structured sentiment JSON

Tool Calling — picks a tool, runs it, answers

ReAct — one action / observation per step, with cited sources

ReAct — fallback to web search

Plan-and-Execute — a request-specific plan, executed and validated

Reflection — draft → critique → revise

Reflexion — reflect → web search → cited revision

Reflexion — iterations + citations

Run locally

git clone https://github.com/maxkangdev/agent-playground
cd agent-playground
cp .env.example .env          # add OPENAI_API_KEY
uv sync
uv run streamlit run streamlit_app.py

Optional keys: TAVILY_API_KEY (web search) · PINECONE_API_KEY (RAG) · LANGSMITH_API_KEY (tracing). Weather, currency, and Wikipedia need no key.

---

How it works & design decisions → ARCHITECTURE.md