Someone spliced Karpathy’s words into an Opus 5.5 tutorial, and the comments aren’t buying it yet
A tweet uses Karpathy’s views on agents to drive traffic to a seven-layer guide for Opus 5.5 harnesses, and someone in the comments points out that the person being quoted has not endorsed it. The ideas are worth a look, but the way they were stitched together deserves a question mark.
This tweet stitches two things together: a judgment about agents supposedly from Karpathy, and a seven-layer setup guide for Opus 5.5 harness engineering. The video is there, and so is the link to the guide. The comments immediately add another piece of information: reply 2 says this is a “Stolen Authority post,” and the linked person has not endorsed this content. You can read the material, but don’t treat a famous name’s quote as a warranty.
The core of Karpathy’s point is clear: the biggest mistake in AI right now is forcing agents to do everything without first understanding the harness engineering around the model. The tweet also says OpenAI made this mistake in 2016, at a cost of five years. It then presents an evolution chain: Opus 5.5 → Prompts → Agents → Graphs → Harness. These are the tweet’s claims; whether they are true and who they should be attributed to need to be considered separately.
What is a harness? It isn’t the model itself. It manages the instructions, tools, permissions, state, and checks around the model. For ordinary users, you can think of the model as the engine, while the harness is the steering wheel, brakes, dashboard, and maintenance log. If you only swap the engine and don’t install brakes, a long-running task will show you the bill halfway through.
The most practical part of the guide isn’t mystical. It breaks long tasks into seven layers: write workspace facts into CLAUDE.md; turn repeated processes into skills; use MCP to connect source materials; use permissions and hooks to limit which files can be changed; have a reviewer subagent come back with evidence; allocate reasoning effort to different kinds of work; and finally define completion conditions with a goal, requiring drafts, checklists, and progress records to be written to disk. Every step needs an observable result. Reading materials should produce source selections, drafting should produce files, and review should produce findings that can be handled one by one.
The guide also notes that the 60% token reduction mentioned in Anthropic’s early tester reports was the result of that particular model experiment. Whether your own harness saves tokens has to be determined by measurements of task completion. Don’t take someone else’s experimental numbers and put them in your own budget spreadsheet.
Under this tweet, CooperCode in reply 7 puts it more directly than the main post: when long tasks fail, it’s often not that the model lost the thread, but that the harness never defined what “done” means. That judgment is worth remembering. Models forget, sessions break, budgets burn. Saving progress, verifying results, and reusing workflows—these three things are closer to engineering than swapping in a louder model name.
The comments aren’t that uniform either. Someone asks whether agents are just a set of instructions. Someone complains that big companies don’t dare use them. Reply 5 posts an image.
Back to the controversy at the beginning. Putting Karpathy’s words at the very front adds a layer of authority to the Opus 5.5 guide that follows. Reply 2’s accusation precisely strips away that layer of authority. When reading threads like this, first separate three things: who said it, who has not endorsed it, and whether the guide itself has actionable details. If the first two have problems, the third can still be worth reading. The guide’s seven-layer setup, evidence review, and progress recovery are all things that can be put into practice. Don’t use a famous person’s headshot as your acceptance criterion.
发布时间: 2026-10-29 21:37