Karpathy says delete prompts and keep only Graph; most people get stuck at the Agent layer

Andrej Karpathy, in a video posted in a tweet, breaks AI workflows into four layers: Prompts, Agents, Loops, and Graphs. The tweet says that in 124 minutes he unpacks how to build a graph and why the graph is the only layer that can hold up. Karpathy's exact words are: Prompting is going away. Delete everything, keep Graph.

This line is crisp, and it is easy to misread as saying prompts are useless. According to the layering in the tweet, a prompt is an instruction the user gives the model; an agent calls tools to get things done; a loop keeps an agent running repeatedly until the task is complete; and a graph determines which agents exist and what each agent hands off to the next. Most people stop at the agent layer and never get to the graph.

Graph here means an agent orchestration graph, not a graph from a graph theory class. It governs control flow: who runs first, who waits for whom, who takes over after a failure, and who receives the output. A loop can keep a single agent spinning until the task ends; a graph has to put multiple loops and multiple sets of tools into the same dependency graph. The slogan is crisp, but in practice implementation usually gets stuck first on permissions, logs, and who signs off.

Prompts have not disappeared into thin air. Under this layered view, they are downgraded from manual input in a chat box to local constraints on graph nodes. What is really moved away is control flow. What prompt a node should use still has to be written; it just no longer carries the responsibility of organizing the entire workflow. This change matters to developers, and its impact on ordinary users is more hidden: behind a single button in an interface, there may be a graph deciding the order of calls.

The tweet also links to an article explaining the main steps to master AI: Best article explaining main steps to master AI. In the same discussion, Matt's reply throws cold water on it: A check only informs you if it could have failed. Fix the test, but count it as a new check, on evidence it hasn’t seen. Otherwise you’ve got a co-author signing off on the work. In plain terms, a check can only tell you that it could have failed; after fixing the test, you have to count it as a new check and run it on evidence the model has not seen, otherwise you are effectively letting a co-author sign off on the work.

This reminder is closer to engineering reality than the graph slogan. Having more nodes does not mean higher quality, and having more loops does not mean automatic reliability. A graph can assign division of labor, but verification still has to be independent. Otherwise the entire graph is just copying the same mistake into more branches.

If you remember only one thing: Graph is control flow, not a new prompt. First figure out who is responsible for verification, then draw the graph.