In June, “loop engineering” got a name. Boris Cherny’s line was: “I don’t prompt Claude anymore. I have loops that are running. They’re the ones that are prompting Claude.” In mid-July, “graph engineering” got a name. On July 18th, someone published the obituary: loop engineering is dead. That is six weeks from birth to funeral, and the death certificate cited a single one-line tweet.
I want to argue two things. Both labels describe real work that you may have already done, and the funeral rests on a category error.
Walk one scenario through three stages. Suppose you want an AI-generated industry briefing every morning.
Stage one: you drive. Each morning you type the prompt. An agent runs its reason-act-observe cycle and hands you the briefing. The only loop here belongs to the agent itself, and running it is the harness’s job, one turn at a time.
Stage two: you replace yourself with a machine. A timer fires the prompt at eight. A check rejects the draft if it is thin or uncited, and sends it back until it passes. Notice what you just designed. You did not design a prompt. You designed the system that does the prompting: trigger, acceptance test, retry, stop condition. That design work is loop engineering. I had been doing it for months with a cron-driven agent on a rented box before the term existed, and so has anyone who ever wired up “run the tests, make the agent fix the failures, repeat until green.”
Stage three: one agent stops being enough. The briefing is mediocre because a single context is researching, writing, and reviewing all at once. So you split the job. Three researchers fan out in parallel. A writer sees only their cleaned notes. A reviewer holds a fresh context and a reject edge that routes bad drafts back. Notice again what you designed: an org chart, with roles, ordering, parallelism, merge points, and failure routes. That design work is graph engineering. And inside every box on that chart, an agent is still running its own loop.
An org chart does not replace employees. A graph does not replace loops. A graph is made of them.
That is the whole rebuttal to the obituary. A graph expresses sequence, branching, fan-out, and cycles. A loop is just an edge that points backward. Saying that graphs make loops obsolete is like saying that shapes make circles obsolete. The sober reviews of this episode reached the same verdict: the label is new, but the technique is not, and graph runtimes have been shipping for two years. What died in six weeks was never the practice. Only the novelty of a hashtag died.
Strip away the theater and one real question survives. It is the only question worth carrying out of the debate: can you name fixed roles in this job and draw arrows between them?
If the answer is yes, meaning the same specialists run in the same order on every run, then draw the graph and take the benefits: parallelism, clean per-role contexts, and control flow you can audit before it runs.
If the answer is no, because the cast changes every time, then keep one loop and let a lead agent hire sub-agents as it goes.
A vertical agent I work on answers yes, and it wraps its loop inside a deterministic graph. My morning briefing answers no, and it stays a single stubborn loop. Neither one should be embarrassed about its answer.
I had been running all three stages before either term had a name. The work was real the whole time. The names are only now catching up, and one of them already has a tombstone.