What Happened

Terminology · LLM Engineering · AI Agents · AI · ExplainerFive terms, February 2021 to July 2026

• 5names, each wrapping the last in BusinessNext’s account
• 3different places the loop sits: inside the harness, a floor above it, or inside each node of a graph
• ~20improvements autoresearch kept from about 700 changes, by Karpathy’s own count
• 12design primitives a public harness list names, the agent loop first among them

Five names, in the order they appeared Each card gives the earliest dated use found, and the words of the person who used it.

• 1 · February 2021

  Prompt Engineering 提示工程

  Governs: how one request is worded.

  • The earliest dated use found is Laria Reynolds and Kyle McDonell’s paper of 15 February 2021: “Prompt engineering for a language model whose input and output are in natural language may be conceived as programming in natural language.”
  • By March 2023 Anthropic was advertising a “Prompt Engineer and Librarian” at $175,000 to $335,000 a year.
  • Anthropic’s own definition: methods for writing and organising a model’s instructions for the best outcome.
  • Not replaced. BusinessNext and Databricks both describe it as a part of what came after.

• 2 · June 2025

  Context Engineering 情境工程

  Governs: what the model can see at each step.

  • Tobi Lütke, 19 June 2025: “the art of providing all the context for the task to be plausibly solvable by the LLM.”
  • Andrej Karpathy, six days later: “the delicate art and science of filling the context window with just the right information for the next step.”
  • Anthropic’s Applied AI team formalised it on 29 September 2025 as “the natural progression of prompt engineering.”

• 3 · February 2026

  Harness Engineering 駕馭工程

  Governs: everything around the model — tools, memory, sandbox, guardrails.

  • The earliest dated use found is OpenAI’s, 11 February 2026: five months building a product with Codex writing every line.
  • Agent = Model + Harness. LangChain’s Vivek Trivedy, March 2026: “If you’re not the model, you’re the harness.”
  • Databricks, June 2026, lists what it holds: tools, memory, a workspace, and guardrails such as approvals.

• 4 · June 2026

  Loop Engineering 迴圈工程

  Governs: the system that prompts the agent, so a person does not have to.

  • Peter Steinberger, 7 June 2026: “You should be designing loops that prompt your agents.” Addy Osmani’s “Loop Engineering” is dated the same day.
  • The Claude Code team’s definition: “agents repeating cycles of work until a stop condition is met.”
  • The idea predates the name: Geoffrey Huntley’s “Ralph” of July 2025 was an agent run in a bare shell loop. Karpathy’s autoresearch is BusinessNext’s example.

• 5 · July 2026

  Graph Engineering 圖結構工程

  Governs: which steps run, in what order, who hands off to whom, and what is recorded.

  • Josh C. Simmons, 4 July 2026: “designing agentic systems as explicit graphs instead of implicit loops.”
  • Three parts: nodes that each do one job, edges that carry the decisions, and a state object checkpointed at every crossing.
  • BusinessNext brought the term to Chinese-language readers on 28 July, naming Simmons in its source line.

How firmly each part stands Documented, then self-reported, then contested, then not established.

Timeline

Everything here that has a date 19 dated entries. One term has no first use that could be found: harness engineering. Prompt engineering’s earliest use found is dated, and is not claimed as the first.

1. 15 Feb 2021Laria Reynolds and Kyle McDonell post a paper on prompting GPT-3 that uses “prompt engineering” — the earliest dated use of the phrase found.
2. 30 Nov 2022OpenAI releases ChatGPT. BusinessNext dates the whole sequence from here.
3. 13 Mar 2023A trade report finds prompt-engineer jobs advertised at up to $335,000 a year, topped by a listing at Anthropic.
4. 19 Jun 2025Tobi Lütke, chief executive of Shopify, posts that he prefers “context engineering” to prompt engineering.
5. 25 Jun 2025Andrej Karpathy agrees, and adds that context engineering is “just one small piece” of a thick layer of software around the model.
6. 14 Jul 2025Geoffrey Huntley publishes “Ralph”: “In its purest form, Ralph is a Bash loop.” The loop, before it had a discipline’s name.
7. 29 Sep 2025Anthropic’s Applied AI team publishes a guide to context engineering for agents.
8. 22 Oct 2025LangGraph, LangChain’s lower-level agent framework and runtime, reaches version 1.0.
9. 11 Feb 2026OpenAI’s Ryan Lopopolo publishes “Harness engineering”, the earliest dated use of the term found.
10. 6 Mar 2026Karpathy’s autoresearch repository is created.
11. 9 Mar 2026Karpathy reports autoresearch’s first round on nanochat.
12. 10 Mar 2026LangChain’s Vivek Trivedy: “Agent = Model + Harness.”
13. 2 Apr 2026Birgitta Böckeler of Thoughtworks publishes harness engineering for coding-agent users, on martinfowler.com.
14. 13 Apr 2026A position paper on arXiv describes the agent loop as a scheduler with exactly one unit ready to run at a time.
15. 7 Jun 2026Peter Steinberger posts that developers should be designing loops, not prompting; Addy Osmani publishes “Loop Engineering” the same day.
16. 17 Jun 2026Databricks explains the agent harness and puts prompt and context engineering inside it.
17. 30 Jun 2026The Claude Code team publishes its taxonomy of loops: turn-based, goal-based, time-based and proactive.
18. 4 Jul 2026Josh C. Simmons publishes “We Are Entering the Graph Engineering Phase”.
19. 28 Jul 2026BusinessNext publishes its five-term explainer, bylined 陳建鈞.

The Argument

What the explainer argues Not replacement but containment — and then a brake on its own claim.

Who puts what inside what Seven named sources. Read down the right-hand column and the stack does not hold still.

What Others Add

The five, term by term Each definition in its author’s own words.

How Simmons says to practise graph engineering Seven practices, in his order. BusinessNext’s three principles are condensed from the second, third, fourth and sixth.

1. 1Draw the state firstWrite down what the system knows at each point in a run before writing a prompt. “If you cannot write down what the system knows at each point in a run, you have a demo.”
2. 2Keep nodes boringOne job each, testable alone, replaceable whole. “A node that does five things is a loop with extra steps.”
3. 3Put the judgement in the edgesSome hand-offs are hard rules, some are the model’s call; default to the rule wherever you can afford it. The model-decided edges are where the system will fail, so watch them.
4. 4Checkpoint at every crossingA failure then means retrying one node, not the run, and a run can wait three days for a human sign-off.
5. 5Treat humans as nodesAn approval gets an edge in and an edge out, designed like any other step — not bolted on as an exception.
6. 6Put the budget in the stateTokens, money and wall-clock time live in the state and are enforced at the edges. If an agent cannot be stopped at a spend threshold, it is not running itself — it is running up a bill.
7. 7Evaluate the path, not just the outputAn output check says the run ended well; a trajectory check says whether it took a sane path at a sane cost.

The loop and the harness, in practice Two projects whose own builders published the numbers.

Conclusion

What to take away The names are dated. The stack is not.