Karpathy's 2-Hour Free Course Is Going Viral, but the Comments Are Arguing About Three Things

Karpathy spent eight years at OpenAI and Tesla. Someone distilled what he learned into a 2-hour lecture, with the title Agents → Harness → Loops → Graphs, and added: some bootcamps charge $15K and teach less than this.

The tweet got passed around a few times, and the comments quickly split into three camps.

The first camp bookmarked it outright. Someone wrote, '2 hours for his 8 years is worth more than the tuition I paid for a whole semester.' The second camp called it a scam, with a straightforward reason: eight years of experience crammed into two hours just doesn't sound right. The third camp was correcting things: the video isn't from 'last week'; the material is at least two years old. Others asked about the YouTube title, and someone offered to make bilingual Chinese-English subtitles.

What really shifted the discussion was the long article linked under the tweet: 'Harness Engineering: The Complete Guide to Building AI Agents That Don't Fall Apart,' by Lunar. The article opens by naming a common reflex: when an agent fails, change the prompt; when it fails again, add instructions; if it still doesn't work, switch models, add tools, expand context.

But many failures aren't reasoning problems; they're environment problems. The agent doesn't know which files matter, uses the right tool in the wrong place, loses last round's decisions, says it's done without running checks, repeats the same action after a partial failure, and permissions aren't locked down. The model itself isn't broken; a piece of the system around the model is missing.

Lunar calls this piece the harness. It doesn't change one attempt; it changes every attempt.

The article makes several judgments worth pulling out.

First, give the agent a task contract, not a vague intention. 'Optimize the onboarding flow' works for a chat, not for autonomy. The contract has to answer: what must be produced, where the scope lies, what must not be touched, what evidence counts as done, and which actions need human approval.

Don't dump all the context in at once. Give the map first, then let the agent pull details on demand. The goal is signal density per token, not context length.

More tools isn't better. Twenty tools equal twenty ways to make mistakes. You need a tool gateway: hide irrelevant tools, validate parameters, restrict paths and domains, add timeouts, make retries idempotent, require confirmation for risky actions, and return evidence instead of a single 'success.'

Completion can't rely on the agent saying 'done.' Run the cheapest deterministic checks first: compiler, schema, checksums, queries, tests. The model handles fuzzy judgment; code handles the pipeline. The model can propose completion, but only the environment can prove it.

Verification should attack the result, not offer a second opinion. The worker builds the strongest solution; the evaluator looks for reasons to reject it. If you ask the agent to 'check again' in the same context, it often keeps the assumption that caused the error.

Failure recovery has to start with classification. Direct retry isn't recovery; it's repetition. A retry must change at least one relevant condition, otherwise you're paying to reproduce the same failure.

The article ends with a layered minimum viable harness: bounded tasks, a readable environment, controlled actions, durable execution, evidence, and recovery and learning. Start by eliminating the failures you actually encounter. Don't jump straight into a multi-agent architecture. Complexity should be earned by observed failures.

Someone in the comments put it more directly: no matter how big your bookmarks folder is, it's useless if you don't actually write a harness.

Original tweet link: https://x.com/0xkichiro/status/2115474764672225585

Long article link: https://x.com/i/article/2096536117130481673