What It Is

Terminology · LLM Engineering · AI Agents · AI · ExplainerThe term and its practice, January 2024 to August 2026

The idea is agreed. The name is the mess. Every source that defines it describes the same three-part structure. What they disagree about is what to call it, who named it, and whether it is new.

• 14days between the careful essay that coined the term and the joke that made it travel
• 12words in the post that did it — a question, claiming nothing
• 1GIF and one full stop: the whole text of the article that declared loop engineering dead
• 2024the year a paper title set flow engineering against prompt engineering — the earliest use of that phrase found
• 0measurements comparing a graph against a loop

• Nodes

  “A good node is boring”

  A unit of capability — and a person counts as one.

  • In Simmons's definition 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, can be tested alone, and can be swapped out without touching anything else.
  • LangChain describes its own framework the same way: a node can be deterministic code, a single model call, a tool call, or a full agent with its own internal loop.

• Edges

  Edges are decisions

  And they are where it will fail.

  • 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, in Simmons's words, is knowing which is which, “and defaulting to deterministic everywhere you can afford to”.
  • Google's argument for its Agent Development Kit 2.0, as Turing Post reports it, is the same one: “code should control predictable routing, while models should handle the steps that require interpretation or judgment.”

• State

  A schema, checkpointed at every crossing

  Not whatever is sitting in the context window.

  • The useful properties fall out of this. Failure stops meaning “restart the run” and starts meaning “retry the node”, and a task can wait three days for a human sign-off without holding a context window hostage.
  • Budget lives in the state too — tokens, money, wall-clock time — 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. The Taiwanese account of the term makes the same point in almost the same words.

• Verification

  It has to come from outside

  Agents checking agents is not enough.

  • 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.
  • He credits Carlos Perez with the warning behind it: a graph of agents checking agents “can produce extremely organized nonsense”, because models built alike, reading the same flawed context, tend to agree with each other.

How firm is each of these Sorted by standing: what the documents themselves show, then what a document only claims, then what dated sources contradict, then this article's own reading.

Timeline

A practice with a name, then a new name, then a joke The viral cycle is sixteen hours long and sits near the end of a chain that begins in 2023. Times on 18 July are UTC.

1. Aug 2023A graph framework for agents begins. LangChain's LangGraph repository is created. Its first release follows on 8 January 2024 — which is how LangChain can call it three years old in July 2026 while others date its launch to January 2024.
2. 16 Jan 2024“Flow engineering”, in a title. A paper from CodiumAI is called From Prompt Engineering to Flow Engineering: instead of one carefully written prompt, a staged, test-driven flow that generates, runs and fixes code. On its own benchmark GPT-4 solves 44% of validation problems with five attempts each, against 19% with a single prompt.
3. Oct 2025LangGraph 1.0. The graph framework most of these sources name reaches its first major version.
4. 18 Jan 2026The practice, named in a survey. A survey of agentic AI carries a section headed Graph based orchestration and flow engineering: production systems model the workflow “as a state machine where nodes represent tool calls or LLM invocations and edges represent permissible transitions. This approach is often described as flow engineering.”
5. 13 Apr 2026The formal case against the loop. A single-author position paper describes an agent loop as “a single ready unit scheduler” whose next step “comes from opaque LLM inference rather than an inspectable policy”, and proposes a Structured Graph Harness. Its abstract ends: not a production implementation or empirical results.
6. 15 Apr 2026Loops, before the word. On stage at Acquired Unplugged in San Francisco, Boris Cherny, who leads Claude Code, describes his work as writing loops — as Addy Osmani later quotes him, “My job is to write loops”. He thanks the hosts on X that evening.
7. 7 Jun 2026Loop engineering is named. Peter Steinberger posts that you should be “designing loops that prompt your agents”; the same day Addy Osmani publishes an essay titled Loop Engineering.
8. 30 Jun 2026A buzzphrase. Andrew Ng calls loop engineering “a hot buzzphrase”; Anthropic's Claude Code team publishes a guide to getting started with loops.
9. 4 Jul 2026The earnest coinage. Josh C. Simmons publishes We Are Entering the Graph Engineering Phase: a definition in three parts, seven practices, and six sources — one of them the January survey, cited for “graph orchestration and flow engineering”. In the same essay: “So the practice converged and the vocabulary didn't. Fine. Here is the vocabulary: graph engineering.”
10. 18 Jul · 00:34Twelve words. Peter Steinberger, on X: “Are we still talking loops or did we shift to graphs yet?” It proposes nothing and reaches 3.1 million views. The first replies: 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”.
11. 18 Jul · 02:10The first attempt at a meaning. Shubham Saboo: “wtf is a graph. Loops made Agent behavior programmable. Graphs make agent orgs programmable.”
12. 18 Jul · 05:09The obituary. Hamel Husain publishes an X Article: Loop Engineering Is Dead. Enter Graph Engineering. Under a cover drawing of a loop opening into a network, the body is a reaction GIF captioned “Stop it.” and a full stop.
13. 18 Jul · 05:16The 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.”
14. 18 Jul · 16:26The slogan. Santiago: “Loop Engineering is dead. Long live Graph Engineering!” Nine minutes later Husain, quoting his article again: “Bookmarked 100 times 🤣🤣”.
15. 20 Jul 2026Is it real? Turing Post reads the founding post as a joke — Peter is having fun — reports courses, roadmaps, tool stacks by Sunday, counts four different things people mean by graph, and lands the line the argument turns on: a loop is already a graph.
16. 22 Jul 2026Said out loud, twice. Louis-François Bouchard: “it was mostly all a big meme around creating new terms” — and “the joke worked because it points at something real”; in a longer version the same day, “both tweets were jokes”. LangChain's founders, also that day, write that graph engineering “isn't a new idea. It's the latest name for a well established approach.”
17. 28 Jul 2026It reaches Taiwan as vocabulary. 《數位時代》 (BusinessNext) runs it as one of five AI engineering terms, credits X and Josh C. Simmons as its sources, discloses that its first draft was written by AI, and says plainly that how the five layers nest has no settled answer.
18. 21 Aug 2026Introduced again, as new. A survey by 35 authors: “we introduce Graph Engineering, a novel engineering paradigm”. It mentions none of the July posts, nor the phrase flow engineering, and already names the next rung: ontology engineering. A revised version follows on 26 August.

The Argument

He had the older name, and chose a new one The coining essay is careful about its papers, wrong about the previous term, and knew the practice already had a name.

The obituary had nothing in it The article that declared loop engineering dead is a headline, a drawing, and a GIF that says “Stop it.”

What Others Add

What you actually build Josh C. Simmons's seven practices, which the Taiwanese account condenses into three. Strip the fortnight away and the advice is specific, and older than the name.

1. 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.”
2. 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.”
3. 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.
4. 4Checkpoint at every edge crossingFailure becomes “retry the node”, not “restart the run”, and a graph can wait three days for a human sign-off.
5. 5Treat humans as nodesAn edge in, an edge out, a person in the middle. Bolting approval on as an exception handler “is how you get systems that technically have oversight and practically have none.”
6. 6Put budget in the stateTokens, money and wall-clock time live in the state object and are enforced at edges.
7. 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 Every ladder has prompt at the bottom and graph near the top. They disagree about harness — and the newest already has a rung above graph.

And still nothing measures it The formal paper disclaims empirical results. The survey that introduces the paradigm says the measurement does not yet exist.

Conclusion

What to take from it The idea is sound. The launch was a joke, and the joke was about naming. Keep the two apart.