LangGraph Added Zero Prompt Tokens vs No Framework

I measured LangGraph overhead against a plain no-framework agent loop, 200 runs each: zero extra prompt tokens and under 20ms of added latency at p95. The OpenAI Agents SDK sent 9.7% more bytes over the wire and was billed for the same 449 tokens. What you'll learn: Build one tool-calling agent three ways (raw loop, LangGraph, Agents SDK) from a single byte-identical contract Put a FastAPI LLM proxy between each agent and the API to capture real payloads and billed tokens Separate framework latency from model latency with a p50/p95 agent framework benchmark Plant one identical bug and time how fast each version leads you to the failing line The build starts from an empty folder on Windows: hand-written tool schemas, two deterministic tools, then three agents pinned to the Chat Completions API so the endpoint never becomes a second variable. A bench script pushes every run through the proxy and compares token usage, request bytes and LangGraph overhead side by side, then the bug test shows which version is easiest to debug. [GitHub Repo] https://github.com/ByteBuilderLabs/Pr... [LangGraph Docs] https://docs.langchain.com/oss/python... [OpenAI Agents SDK Docs] https://developers.openai.com/api/doc... [OpenAI Chat Completions API] https://developers.openai.com/api/ref... Subscribe to ByteBuilder for AI engineering tutorials that go from concept to working code. [GitHub] https://github.com/ByteBuilderLabs