'Prompting is over': a study path through harness, loop and graph
Nov 2, 2026Original Andrej Karpathy · autoresearch (GitHub) ↗
In 3 lines
- The weight is moving from writing prompts to designing the environment, checks and division of labor an AI works in.
- A seven-level ladder: harness (tools, memory), loop (a check that can fail), graph (split work, gates).
- Study Karpathy's seven repos, then autoresearch, then Anthropic's review post, in that order.
Last week posts on X invoking Karpathy and Anthropic claimed prompting is over and the graph is what remains, drawing hundreds of thousands of views. This piece drops claims with unclear video or quote sources and sets out a study order from the verifiable harness, loop and graph framework and Karpathy's repositories.
The viral claim that 'prompting is over': how much holds up
Posts last week said prompting is fading and only 'the graph' survives, attributing it to a Karpathy lecture or an Anthropic head of research who supposedly said 99% of engineers run 300+ agents. No original source for those quotes or videos turned up. One reply said a clip was Karpathy's February 2025 'How I Use LLMs'. What does hold up is below: the weight moves from writing a good question to designing the environment and procedure an AI works in.
X 게시글·댓글 · x.com ↗Level 1, harness: everything left once you remove the model
X developer m4lka wrote that people who use agents often become the agent's loop: it writes, you read; it says done, you check; it breaks things, you explain. He split the job into harness, loop and graph. The harness is what remains on the architecture diagram once you delete the model: shell, files, test runner, git, and a workspace. More tools is not better; ten well-documented tools beat forty overlapping ones.
m4lka (X 아티클) · x.com ↗The harness's second job: memory and walls
When the context window fills, an agent forgets fixes it already tried and repeats them. The post advises writing the goal, attempts and next step into a file such as PROGRESS.md: the context window is RAM, the harness is disk. 'Please don't touch it' is a wish; a permission the agent does not have is a wall. Examples are an isolated workspace, an allow-list of commands, and time and token budgets.
m4lka (X 아티클) · x.com ↗Level 2, loop: one check that can fail
A loop is produce, check, correct, repeat, and the author says only the check matters. Passing tests, a trace showing the test really ran, or a reviewer in a separate context are strong evidence; the model saying it is confident is the weakest. Write a test that reproduces the bug before the fix and watch it fail once. If you cannot write the check, he adds, you do not yet understand the task.
m4lka (X 아티클) · x.com ↗Level 3, graph: split the work into nodes and edges
Running everything in one line makes the whole slow and single-minded even if each step is right. So split the work into nodes (one job each) and edges (real dependencies). For every arrow, ask what would break if B started before A finished. If nothing, the arrow is a habit and the steps can run in parallel. Merging and sorting should be code, not a model. Here 'graph' means a workflow graph, not an AI knowledge graph.
m4lka (X 아티클) · x.com ↗Gates priced by the cost of undo, and mistakes turned into rules
Gates depend on what it costs to take a change back, not on the model's confidence. A docs change reverted by one commit can ship on green; migrations, deletions and sent emails need a person's sign-off. Level 7 means the same class of failure never costs twice: a missing log view becomes a read-only tool, overlapping work becomes a split rule, each on the layer that owns it. The author says most tasks finish at level 4.
m4lka (X 아티클) · x.com ↗Karpathy's loop in the flesh: autoresearch
Karpathy's autoresearch repository lets an agent edit training code, train for five minutes, keep or discard the result, and repeat. Its README estimates about 12 experiments an hour and about 100 while you sleep. Fortune reported a run of 700 experiments over two days that found 20 optimizations, speeding training by 11% on a larger model. The setup is one file, one objective metric and a fixed time per experiment: the 'check that can fail' from the post above.
How the Claude Code team does code review
In a Claude Code team post shared by Addy Osmani, Claude writes about 90% of the code. Merged PRs more than doubled from June to September while the share of reverted PRs fell by over 40%. Daily merges in the monorepo went from about 2,500 in May to over 30,000 in September. The method was not less review: runnable verification before the PR, a fresh-context subagent reading the diff, and human approval scaled by risk, like the gates above.
Addy Osmani (X) · x.com ↗Study order 1: Karpathy's seven repositories, bottom to top
The order from X user adon: micrograd (backprop in about 100 lines of Python), nn-zero-to-hero (lectures from one neuron to GPT), minbpe (tokenization), nanoGPT (training a GPT), llm.c (training in C/CUDA), nanochat (a whole chat model), autoresearch (an agent that researches overnight). Stars when checked were about 19k, 25k, 11k, 64k, 31k, 60k and 99k, around 310k total. 'A free AI degree' is hype; it is a set of repositories, and nn-zero-to-hero is eight lectures. To learn how models are made, start with the first four; for agents, start at the last.
adon (X) · GitHub · github.com ↗Study order 2: for developers who use agents
Three steps. First read m4lka's harness, loop and graph post and find your symptom from the top: copy-pasting means level 1, checking every 'done' by hand means level 3. Then read the autoresearch README and move its one-file, one-metric, time-limit structure to your own work. Finally compare the verification steps and risk-based approvals in the Claude Code review post. TechCrunch reported Karpathy joined Anthropic's pretraining team on May 19 this year.
TechCrunch · techcrunch.com ↗
