Thanks to everyone who came today! I shared a lot of information, so here is the deck so you can follow along!
Every agent framework runs your agent for you, up to a point. Each one handles some of what an agent needs, then leaves the rest for you to write yourself or pay another tool to cover. I call that leftover work. It decides how much time and money a framework costs you once your agent runs every day.
In this session, I compare four ways to build an agent and show you what each one leaves on your plate:
Writing the code yourself. Full control and easy model switching, but you build pausing, restart recovery, and everything else from scratch.
The SDKs from the AI labs. Claude Agent SDK, OpenAI Agents SDK, and Google’s ADK. They write the loop for you, though they tie you to one lab’s models.
LangGraph. Pausing and restart recovery come built in once you connect a database.
n8n. A visual tool where you drag steps onto a workspace, with limits that catch up to you as your agent grows.
I also clear up the four products with confusingly similar names: LangChain, LangGraph, LangSmith, and Langfuse.
Then I get to the one requirement none of the four handles: when a server crashes at the wrong moment, an agent can repeat a step and write the same record twice. I walk through why it happens and the fix that stops it, called an idempotency key.
This Sunday, I put every claim from today to the test. I’ll build the same agent four separate times, once in each framework, then break all four on purpose. I crash them right after they write to a database, restart the servers mid-run, change the code while a run is paused, and swap out the model. You get to see which tools hold up and which fall apart, plus the code to run these tests on any tool you’re weighing before you commit.
