Stop staring at every step: let the AI agent run itself; you only approve the last step
A 70-minute slide-free Karpathy demo and a long essay about loops and graphs are both saying the same thing: don't spend your energy staring at every step; get the structure right, and just give the final nod.
In a widely circulated screen recording, Andrej Karpathy opens up his screen and demonstrates a real 70-minute workflow. No slides, no opening remarks—just Prompt, Context, Tools, and Workflow. After watching it, you'll understand that the gap between people using AI lies in how they use it, not in the tools.
Karpathy's approach is simple: choose the right model, feed it the right context, and leave the rest to tools. Anyone can open ChatGPT and type, but almost no one organizes a workflow the way he does.
Below the video, Hanako wrote a long essay deconstructing the same thing from an engineering perspective. Its core is two words: loop and graph.
A loop is a unit that can self-correct. It consists of four steps: generate, check, correct, and repeat until it turns green. Checking is the soul of the entire loop. Hanako says you should write the check conditions first, in a form a program can evaluate. For example, the test suite exit code is 0, every conclusion has a source line, and the diff only touches files within the plan. But 'the output looks good,' 'the model is confident,' and 'there are no errors' do not count as checks. No errors does not mean it is correct.
The boundary of a loop is also clear: it can only make a single unit better. It doesn't know which units should exist, nor does it know the order. So you end up with a very hardworking agent running the wrong three steps in the wrong order. Every step is right, but the result is wrong. At that point, the problem is usually in the structure of the graph.
A graph is the shape of a workflow. Nodes are units; edges are dependencies. The most common mistake is treating 'then' as an edge. There is no edge between 'summarize the file' and 'check the weather,' because the weather does not consume the summary. If you can't name the variable that crosses an edge, then there is no edge, and waiting is just waiting in vain.
There are four kinds of nodes in a graph: a splitter divides the work, a worker executes it, a code node handles fixed operations, and a gate grants passage. The splitter matters most: split it wrong, and everything downstream is wasted. The code node is the easiest to overlook: merging, deduplicating, sorting, comparing—a few lines of code are enough for these. Calling a model only adds cost, latency, and randomness. Any operation that can be described in code should not be left to model inference.
Loops live inside nodes; graphs live between nodes. A loop without a graph is single-point optimization; a graph without a loop is parallel fault generation.
A graph also needs two return paths. One is the correction edge, which is short: the gate sends a unit back to fix the current run. The other is the learning edge, which is long: accepted results go back to the splitter and become constraints for how to split work later. Most people build only the correction edge, so the system runs fast but never gets smarter. The learning edge carries constraints, not outputs.
When sending work back, return only the failed unit, not the entire batch. Suppose four slice transplants are done and one fails; sending all four back means three correct ones get rewritten, and then everything is validated again. One failure turns into four uncertainties. The return message must clearly define scope: fix only the specified files, and don't casually change anything else. Limit retries to three; three failures mean the problem is the plan, and the loop cannot see the plan.
Whether a gate opens depends on the blast radius. Confidence is not trustworthy, because it is the only variable the model can influence. If undoing is cheap, you can let it through first; if undoing is expensive, hold it with deterministic checks; for irreversible changes—such as migrations, deleting data, or moving money—that gate never opens. Thresholds get adjusted; closed gates do not.
Karpathy's video and Hanako's essay are saying the same thing: draw the structure correctly, let the right things happen automatically, and keep humans only for the final step. Someone commented that this 70-minute work recording is the only course needed, and everything else is commentary. Others say the video is outdated. But the core hasn't changed: the value of AI depends on the structure you build; the number of prompts doesn't matter.
Most people will keep tuning a loop and call it a system. A few will draw the graph and let loops run inside it. The difference is in the structure, not in the effort.
发布时间: 2026-11-22 14:08