Karpathy Says "Prompting Is Dying"; What Really Remains Is Only the Graph
Andrej Karpathy said in a two-hour demo: Prompting is about to disappear; the only thing left is the graph. m4lka explains across seven levels why agent systems are such that 'the prompt is the entry point, the graph is the backbone,' and includes a self-check checklist.
Andrej Karpathy (founding member of OpenAI and former Director of AI at Tesla) dropped a line in a two-hour demo: "Prompting is going away. Delete everything and keep the graph."
This video spread quickly on X. The sharer's comment was even more blunt: a loop keeps an agent running until the task is complete; a graph determines which agents exist and what each agent passes to the next. Anyone can build an agent, but almost no one builds the graph that runs them.
m4lka published a long article in October titled "Harness, Loops, Graphs." The opening line is blunt in a way that stings a little: "Your agent is not autonomous; you are its loop." You inspect every step because there is nothing else in the system that can do it.
The article divides agent systems into three roles: the harness determines what an agent can do, the loop determines what is true, and the graph determines what happens next. Do not treat them as three parallel tools. They are a ladder, and each layer exists because of a specific failure in the next layer.
Seven levels, each solving one failure and creating the next:
Level 0 Chat: You paste in a bug, the model offers suggestions, and you manually run tests and paste back error messages. The model is just a function; you are the program around it. One person is doing the harness, loop, and graph all at once.
Level 1 Harness: hands: Give the model a shell, file access, and a test runner. It can now act, but it has no memory or boundaries. After 40 minutes, the context is stuffed with test logs, it forgets what it has tried, and it can only try again.
Level 2 Harness: memory + walls: Add PROGRESS.md and git commits, and put state outside the context. Add permissions and sandboxes. A wall is not an instruction. The model not having permission is the wall. It says "Fixed it; all tests should pass."—"should" shows there is no mechanism guaranteeing it.
Level 3 Loop: the check: A check that can fail. Tests first reproduce the problem, then verify the fix. If the model can pass by saying "done," then no matter how long the loop is, it is just repetition.
Level 4 Loop: feedback: Turn "wrong" into "wrong here, which direction to go, and which paths are already dead." There also need to be exit conditions: stop escalating if the same result appears twice; stop when the budget is spent.
Level 5 Graph: shape: Break the task into nodes and keep only real dependencies. If B does not need to wait for A, then the arrow is a habit, not a dependency. A bug-fix task can be split into fix, audit, and documentation running in parallel, then converge at the end.
Level 6 Graph: gates: Nodes that pass the gate are frozen; on failure, only the failed node is rerun. Grade by undo cost: automatically ship what can be rolled back with one click; require human sign-off for what cannot be undone. Focus attention only on the places that cannot be undone.
Level 7 Learning: Write lessons back into the corresponding layer. Failures become new tools, new checks, and new decomposition rules. The same mistake should not cost you twice.
This ladder diagram breaks "automatic agents" into diagnosable steps. Karpathy's "keep the graph" can land within this framework: a prompt is a one-off conversation; a graph is the backbone of the system. An agent that can fix bugs is not hard; what is hard is letting fixes, audits, and documentation run in parallel without contaminating one another, with every node independently verifiable.
m4lka also provides a self-check list. If you are still copy-pasting between chat and terminal, you are stuck at Level 1; if the agent is retrying methods it has already tried, you are stuck at Level 2; if "done" only counts once you have checked it, you are stuck at Level 3; if retries burn budget on the same error, you are stuck at Level 4; if every step is correct but the whole is slow, you are stuck at Level 5; if one failure causes all work to be redone and you have to look over everything, you are stuck at Level 6; if the same error shows up once a week, you are stuck at Level 7.
You cannot skip levels. Ten agents without checks are ten times the cleanup, not ten times the output. Design starts with the check at Level 3: if you cannot articulate "what counts as green," no layer after that can save you. Most tasks only need to reach Level 4; before climbing higher, calculate the cost of failure.
The full article is here.
发布时间: 2026-11-20 03:41