What It Is
LLM Engineering · Terminology · Source Verification · AI · ExplainerTratopedia · 16 Aug 2026
Fifth of five · the one that arrived twice
One Term, Coined Twice, Two Weeks Apart. The Joke Won.
On 4 July 2026 Josh C. Simmons published a careful, sourced essay coining graph engineering. It sank. Fourteen days later, at 00:34 UTC on 18 July, Peter Steinberger asked twelve words on X — Are we still talking loops or did we shift to graphs yet? — and drew 3.1 million views. Four hours and thirty-five minutes after that, Hamel Husain published Loop Engineering Is Dead. Enter Graph Engineering, behind a paywall almost nobody could get through. Both posts were jokes, and a named analyst said so four days later. Within forty-eight hours the term had courses, roadmaps and three competing definitions. And a survey published in January had already given the practice a name: flow engineering. Nobody noticed that either.
The idea is agreed. It is the name that is a mess. Four independent sources describe the same three-part structure in almost the same words. Then they disagree about what to call the layer, and about whose ladder it sits on.
- 14days between the earnest coinage and the joke that made the term travel
- 12words in the post that did it — phrased as a question, claiming nothing
- 6months earlier a survey had already named the practice flow engineering
- 0measurements comparing a graph against a loop. Two terms running now with none
Nodes
“A good node is boring”
- A node can be a model running the familiar think-act-observe cycle, a plain deterministic function, a retrieval step, or a human being. It does one thing, you can test it alone, and you can swap it out without touching anything else.
- Treating approval as a node rather than as an exception handler is the difference between a system that has oversight and one that only technically has it.
Edges
Edges are decisions
- A typed transition carrying state. Some are deterministic — tests pass, deploy. Some are model-decided — does this ticket go to billing or to abuse. The job is knowing which is which, “and defaulting to deterministic everywhere you can afford to”.
- Three sources reach this independently, and all three add the same warning: the model-decided edges are where the system goes wrong, so instrument them.
State
A schema, checkpointed at every crossing
- This is the commitment the useful properties fall out of. Failure stops meaning “restart the run” and starts meaning “retry the node”; a task can wait three days for a human sign-off without holding a context window hostage; and there is an audit trail that is not somebody scrolling a transcript.
- Budget lives in the state too — tokens, money, wall-clock — and is enforced at edges. If you cannot stop an agent at a spend threshold, you are not running an autonomous system. You are running up a bill.
Verification
It has to be external
- Louis-François Bouchard: Asking a model to review its own answer is still useful, but it is not external verification. A test runner, execution result, transaction, customer outcome, or expert human review gives the system evidence from outside its own reasoning.
- This is the one commitment that carries over unchanged from the previous article in this series, and the one every source states in some form.
How firm is each of these One row cannot be graded because the document behind it cannot be read. That is not a footnote — it is the reason this article exists.
| How firm | What | On what |
|---|---|---|
| Read at source | The definition in three parts — boring nodes, typed edges, checkpointed state — and the loop is not dead, it got demoted | Josh C. Simmons, drjoshcsimmons.com, 4 Jul 2026 |
| Read at source | The twelve words, and 3.1M views. The replies are all jokes — im tired boss, bro stop I'm on vacation | Peter Steinberger on X, 18 Jul 2026 00:34 UTC |
| Title and date only | Loop Engineering Is Dead. Enter Graph Engineering exists, and is timed 4h35m after the post above. Its body is behind X Premium and this article has not read it, so nothing is said about what it argues | Hamel Husain on X, 18 Jul 2026 05:09 UTC |
| Read at source | “A loop is already a graph. It is simply a graph whose path returns to an earlier node.” | Ksenia Se, Turing Post, 20 Jul 2026 |
| Read at source | “It was mostly all a big meme around creating new terms” — and, in the next breath, “the joke worked because it points at something real” | Louis-François Bouchard, 22 Jul 2026 |
| Read at source | “This approach is often described as flow engineering” — nodes, edges, state machine, LangGraph, six months early | arXiv:2601.12560 §5.1, submitted 18 Jan 2026 |
| Read at source | The scheduler framing — a ready set of exactly one — and the paper's own disclaimer: not a production implementation or empirical results | Hu Wei, arXiv:2604.11378, submitted 13 Apr 2026 |
| Computed, not read | Every timestamp on 18 July, to the second. X post IDs are snowflakes, so (id >> 22) + 1288834974657 gives the creation time — which is why this article can say 18 July where others say 17 | Derived by this project from the post IDs |
| This project's reading | That the term has two unrelated origins, and the comic one is the one that travelled | Marked as a reading, not a sourced fact |
Timeline
January names it. July discovers it. The viral cycle is sixteen hours long and sits at the bottom of a chain that started six months earlier. Times are UTC.
- 18 Jan 2026The practice already has a name, and it is not this one. A survey of agentic AI carries a section headed Graph based orchestration and flow engineering: “A key industry shift is the move from open ended multi agent chat loops toward explicit workflow graphs… This approach is often described as flow engineering.” Nodes, edges, a state machine, LangGraph as the example. Everything July would announce.
- 13 Apr 2026The formal case against the loop. A single-author position paper characterises an agent loop as “a single ready unit scheduler: at any moment, at most one executable unit is active, and the choice of which unit to activate comes from opaque LLM inference rather than an inspectable policy.” It proposes a Structured Graph Harness — and says of itself, in the abstract, not a production implementation or empirical results.
- 4 Jul 2026The earnest coinage. Josh C. Simmons publishes We Are Entering the Graph Engineering Phase: seven minutes long, six citations, a definition in three parts, and practice advice. In the same essay he writes “let me name it retroactively: loop engineering” — a term already four weeks old and in wide use. The essay sinks.
- 18 Jul · 00:34Twelve words. Peter Steinberger, on X: “Are we still talking loops or did we shift to graphs yet?” It is a question, it proposes nothing, and it will reach 3.1 million views. The replies understand it perfectly: im tired boss; bro stop I'm on vacation; YouTube for the next 3 weeks “WHY YOU NEED TO BE RUNNNG A GRAPH NOT A LOOP”.
- 18 Jul · 05:09The obituary, four hours and thirty-five minutes later. Hamel Husain publishes an X Article: Loop Engineering Is Dead. Enter Graph Engineering. It sits behind X Premium. This article has not read it and does not say what is in it — and neither, at the time, could most of the people repeating it.
- 18 Jul · 05:16Six minutes later, the wink. Husain, quoting his own article: “If I saw this on the TL I would personally be afraid to click on it. But I promise its not what you think it is.” By the evening Santiago has posted Loop Engineering is dead. Long live Graph Engineering!
- 20 Jul 2026Two days in, someone asks whether it is real. Turing Post: Is Graph Engineering Real? It reads the founding post as a joke — Peter is having fun — notes that by Sunday the timeline had produced courses, roadmaps, tool stacks, and lands the sentence the whole argument turns on: a loop is already a graph.
- 22 Jul 2026Four days in, a named analyst says it out loud. Louis-François Bouchard: Steinberger posted one tongue-in-cheek line, Husain played along, “and suddenly, we had a new discipline. But it was mostly all a big meme around creating new terms.” Then he refuses to stop there — the joke worked because it points at something real — and gives the five-rung ladder, harness included.
- 28 Jul 2026It reaches Taiwan as settled vocabulary. 《數位時代》 runs it as one of five AI engineering terms, crediting X、Josh C. Simmons in the footer without linking him — and disclosing that the piece's first draft was written by AI. It is the article that set this five-part series going, and it says plainly that the ordering of the five has 目前還沒有定論.
The Argument
The man who named it got the previous name wrong Not a gotcha. It is the most useful thing in the story, because of what it says about how fast this vocabulary moves.
Josh C. Simmons's essay is the careful one. Two arXiv identifiers, both real; a survey figure of 60 per cent, quoted accurately; a characterisation of a paper's section that turns out to match its actual section heading almost word for word. This project went looking for overclaiming and did not find it.
And in the middle of that essay is this:
“Nobody named the discipline, because when there is only one way to build something, the way doesn't need a name. So let me name it retroactively: loop engineering.”
Loop engineering had been named on 7 June, four weeks earlier, in an essay by Addy Osmani that this project covered in the previous article. Boris Cherny had said my job is to write loops on 2 June. Anthropic published an explainer on 30 June. Andrew Ng described three of them on 1 July. Turing Post, writing independently a fortnight later, dates the term to about six weeks old and names Osmani and the same June date.
Nothing suggests bad faith and this article alleges none. The reading it offers is duller and more useful: a writer who verifies his papers and misses a month of X is telling you where this vocabulary actually lives. It does not live in the literature. It lives on a timeline, it moves in days, and a diligent person reading the primary sources can be a month behind without noticing.
Which is the same mechanism that then buried his own essay. He published on 4 July with citations and a definition. Fourteen days later a twelve-word question with no definition in it took the term to three million people. The careful version and the famous version are not the same document, and they never met.
Nobody could read the founding document Four hours and thirty-five minutes after a joke, behind a paywall. Then three competing definitions in two days.
The post that made graph engineering travel is a question. The article that declared the previous term dead is unreadable without a subscription, and this project can confirm that firsthand: fetching it returns a title, a byline, a date, and then a login wall where the text should be. Bouchard notes the same thing — almost nobody could check what it said.
So consider what an interested developer actually had, on 18 July 2026, to work out what graph engineering meant. A twelve-word question. A headline. Its author saying it's not what you think it is. And, if they searched, a fortnight-old essay by someone they had never heard of.
Within forty-eight hours the term had three competing definitions, plus courses, roadmaps and tool stacks, plus threads explaining that Microsoft, Stanford and Anthropic had all discovered graph engineering at once. That is not a coincidence and it is not stupidity. When the founding text cannot be read, every reader supplies their own, and there is nothing to check any of them against.
This is the mechanism the previous article in this series was groping at and did not have a clean case for. Loop engineering never converged because six people answered different questions. Graph engineering had the opposite problem: the technical content was agreed from the start — nodes, edges, checkpointed state, external verification, four independent sources in near-identical words — and the name still arrived attached to a punchline nobody could open.
What Others Add
What you actually build Strip the fortnight away and the advice underneath is specific, agreed, and older than the name.
- 1Draw the state before you write a promptThe state schema is the architecture. “If you cannot write down what the system knows at each point in a run, you have a demo.” Not a system — a demo.
- 2Keep the nodes boringOne thing each, so it can be tested, cached, retried and replaced. “A node that does five things is a loop with extra steps.” Taiwan's 《數位時代》 reached the same rule independently: a step with five things in it is just the old way again.
- 3Put the judgement in the edges, and instrument themThe intelligence of a graph system lives in its routing, so the model-decided edges are exactly where it will fail. Google's ADK 2.0 argument is the same one: code should control predictable routing, while models should handle the steps that require interpretation or judgment.
- 4Checkpoint at every edge crossingFailure stops meaning “restart the run” and starts meaning “retry the node”. This is also what makes long work real: a graph can wait three days for a human sign-off without holding a context window hostage the whole time.
- 5Treat humans as nodesAn edge in, an edge out, a person in the middle. “Bolting it onto the outside as an exception handler is how you get systems that technically have oversight and practically have none.”
- 6Put budget in the stateTokens, money and wall-clock time live in the state object and are enforced at edges. Two sources, writing a continent apart, land on nearly the same sentence: if you cannot stop it at a spend ceiling, it is not a system that runs itself — it is just burning money.
- 7Evaluate trajectories, not just outputs“Output evals tell you the run ended well. Trajectory evals tell you whether it took a sane path at a sane cost, and the path is where your next failure is hiding.”
Whose ladder, and how many rungs The four-rung version drops harness engineering — while leaning on a paper that calls its own proposal a Structured Graph Harness.
| Whose | The ladder | Harness? |
|---|---|---|
| Josh C. Simmons · 4 Jul | prompt → context → loop → graph | Absent. Four rungs, “each a step up in abstraction… less whispering and more architecture” |
| Louis-François Bouchard · 22 Jul | prompt → context → harness → loop → graph | Present. And they “overlap heavily, even if they are not identical” |
| 《數位時代》 · 28 Jul | prompt → context → harness → loop → graph | Present — and it says outright that which layer contains which has 目前還沒有定論 |
| Hu Wei, arXiv · 13 Apr | No ladder — a design proposal | It names its own proposal “SGH — Structured Graph Harness”. The word left off the four-rung ladder is in the title of the design that ladder cites |
The rebuttals are better than the announcement, and two of them come from inside it.
Turing Post supplies the one that does the most work in the fewest words: “A loop is already a graph. It is simply a graph whose path returns to an earlier node.” If that is true — and it is — then the 18 July framing of loops being replaced by graphs was never a technical claim at all. It was a claim about which word people should say.
The second rebuttal is Simmons's own, published two weeks before the movement he is credited with starting: the loop is not dead, it got demoted. Inside a node, a model still runs the same loop it always ran. “Loop engineering was the craft of what happens inside one context window. Graph engineering is the craft of what happens between them.” That is a clean, useful distinction, and it is the first thing the viral cycle discarded.
Then there is the age of it. Turing Post: state machines, DAGs, workflow engines and orchestration systems “have been doing versions of it for decades”, and LangChain shipped LangGraph in January 2024. Simmons himself finds it funny — the frontier is “drifting back toward the oldest discipline in distributed computing, and half the people arriving there are convinced they just invented it.”
And one warning worth carrying: “graph” now means two different things. A control graph is the structure of a workflow. A knowledge graph is information stored as entities and relationships, which is what GraphRAG is about. They are being conflated, and only one of them is what any of this is about.
And still nothing measures it Two terms in a row. The one paper that frames the question rigorously says so about itself, in its own abstract.
The previous article in this series ended on the absence of any measurement holding the agent fixed and varying the loop. The same hole is here, and this time the field's own best paper states it plainly. arXiv:2604.11378 closes its abstract with: “This is a position paper and design proposal. We provide a theoretical framework, design analysis, and experimental protocol, not a production implementation or empirical results.” Its comparison table rates its own proposal “Design only; no empirical validation” against LangGraph's “Widely deployed; community-validated”, under a heading that reads Fairness disclaimer. It is the most honest document in this corpus.
Which makes the one number that did travel worth checking, and it does not survive. The paper's body says “70 open-source LLM agent projects reveals that 60% (42 out of 70) adopt the Agent Loop pattern”. Its appendix prints the complete list with GitHub URLs. Counting them: the Agent Loop heading claims 41, the body says 42, the URLs actually listed come to 49, and the five category headings sum to 68 against a survey stated as 70. Four numbers for one survey.
Four of the five categories reconcile exactly, which is why this is reported rather than blamed on a bad reading. Simmons quoted the 60 per cent figure accurately — the source is what wobbles. And a benign explanation may well exist: a revision, a miscount, a category boundary read wrongly here. What cannot be said is that graph engineering rests on measurement, because the single empirical claim underneath it does not add up to itself.
Conclusion
What to take from it The idea is sound. The launch was a joke. Those two facts are separable, and keeping them separate is the whole job.
Judge the idea, not the launch
It would be easy to read “the founding posts were jokes” as “the term is worthless”. Bouchard explicitly refuses that step and so does this article. Four independent sources — a personal essay, a weekly newsletter, an analyst's substack and a Taiwanese trade magazine — describe the same nodes, the same typed edges, the same checkpointed state, and three of them independently arrive at one job per node and put the budget in the state. That is convergence, and it is exactly what the previous term never managed. The joke and the substance are separable.
An unreadable founding text is a licence to invent
Three competing definitions inside forty-eight hours is not a story about careless developers. The post that made the term famous was a question. The article that declared the old term dead was behind a paywall. With nothing to check against, every reader wrote their own version, and every version was equally unfalsifiable. This project's reading, marked as a reading: that is why it fragmented, and it is a mechanism worth recognising the next time a term arrives this way.
It had a name in January. That is the finding.
A survey submitted on 18 January 2026 carries a section headed Graph based orchestration and flow engineering, opens it with “a key industry shift”, describes nodes and edges and a state machine, names LangGraph, and reports that the approach “is often described as flow engineering”. Six months before the coinage. Not one source in this corpus mentions it — including the coinage essay, which cites that very paper for something else. The thing did not need naming in July. It needed noticing.
Two terms running, and still no measurement
Loop engineering ended with nothing comparing a loop against prompting. Graph engineering ends with nothing comparing a graph against a loop — and with its most rigorous paper disclaiming empirical results in its own abstract, and its one travelling number failing to reconcile with its own appendix. Five terms into this series, the pattern is that the vocabulary moves far faster than the evidence for it, and that the people naming things are frequently the first to say so.