If Nobody Is Raw Prompting Anymore, Why Are Agent Engineering Roles Still Paying $900k?

v3lkera posted a tweet relaying what an Anthropic engineer said: 95% of engineers run self-improving harnesses on Opus 5.5, and in two months nobody will raw prompt anymore. The tweet also links a 35-minute lecture, moving from Prompt to Harness, Loop, Self-Check, and finally to an agent that can improve overnight. The video is below.

Video from v3lkera's tweet: https://x.com/v3lkera/status/2115459815128338659

That sentence is most easily read as prompt engineering being over. The reply from pepegachad under the tweet is more striking: if in two months nobody raw prompts, why are these offers still priced as scarce skills? $500,000 to $900,000. That question is more useful than the slogan.

m4lka's article 'Harness, Loops, Graphs: What Each Layer Fixes and What It Breaks' breaks agent systems into 7 layers. First, some background: the harness is the runtime environment outside the model, including shell, files, tests, git, browser, permissions, and memory; the loop is produce, check, correct, repeat; the graph is breaking a task into nodes and dependencies. Raw prompting means a human directly writes prompts, copies results, runs tests, and pastes errors back in.

m4lka arranges them as a ladder. Each layer exists because the layer below it breaks in predictable ways.

Level 0: Chat. A human pastes a bug report, the model gives a fix, the human copies it into the editor, runs tests, and pastes the error back. The model is a smart function; the human carries the whole program. Fine for one-off problems, but repeat it twice and it's a disaster.

Level 1: Harness. Give the agent a shell, files, tests, git, and a browser. A clean test: remove the model from the architecture diagram, and everything left is the harness. The more tools there are, the more options the model has to choose wrong. Ten narrow, clear tools are better than forty overlapping tools.

Level 2: Memory and walls. Memory goes outside the model, git commits serve as checkpoints, and artifacts are passed by reference. The context window is RAM; the harness is disk. Walls are permissions, isolated worktrees, command allow-lists, and token and time budgets. Please do not modify migrations is a wish; the agent not having permission to do so is a wall.

Level 3: Loop. Four actions: produce, check, correct, repeat. The focus is on check. A check must be able to fail; watch it go red once first. If you can't write a check, the task isn't understood yet. Writing a check is writing a spec.

Level 4: Feedback must converge. A red light only gives a verdict. Feedback should say what went wrong, in what way, and what you've already tried. EXPECTED and SAW turn the red light into direction; FENCE restricts the next attempt to one file; DEAD ENDS prevents attempt 3 from quietly becoming attempt 1. Loops need exits: if the same result appears twice, continuing to try is pointless.

Level 5: Graph. Ask of every arrow: what breaks if B starts before A finishes? If the answer is nothing, that arrow is a habit. Split off the parts that can fail independently. Each worker gets its own room; four agents sharing one context is one agent wearing four hats. Questions with the same answer go to code: merging, sorting, deduplicating, counting. Gates go at node exits.

Level 6: Graph gates and rollback. Freeze outputs that have passed a gate; downstream cannot reopen them. On failure, rerun only the failed node. Approval depends on rollback cost; the model's self-reported confidence comes last after the evidence. The model can mark touches auth; code decides whether that means one approval or two. Humans appear only where errors cannot be undone.

Level 7: System learning. Failures belong to a layer, and lessons should land in the layer that owns them. Write lessons as constraints, not as last week's log dump. Test: the same kind of failure shouldn't cost you twice.

m4lka's conclusion: the harness gives the agent capability, the loop makes the agent honest, the graph makes the organization orderly, and learning makes it better than it was yesterday. Most tasks only need Level 4; climb layers by checking whether the current cost of failure is higher than the cost of building the next layer.

Back to that tweet. 'In two months nobody will raw prompt' does not by itself explain the $500k to $900k offers. The 35-minute lecture in the tweet and m4lka's seven-layer framework point to the same thing: the money goes to people who can build the harness, loop, graph, and rollback permissions. Raw prompting will disappear; system design will not.

m4lka original article: https://x.com/i/article/2114299133674565632

Quoted tweet: https://x.com/m4lkxx/status/2114338456750997922