What Happened
LLM Engineering · Source Verification · AI · Software Engineering · ExplainerBusinessNext · 28 Jul 2026
Source review · five names, checked one at a time
The Picture Is Surer Than the Text
BusinessNext introduced five kinds of “AI engineering” in a graphic and an article: prompt, context, harness, loop, graph. Both say the same thing about how they relate — not new replacing old, but each one wrapping the last, the older becoming a component of the newer. What differs is how sure they sound. The graphic dates every layer and draws the nesting as settled. The article dates two things, then says outright that the industry has not agreed which layer contains which. A reader who sees only the graphic would not know that. Two sources here were read at their own publishers; everything else, including the article's own web page, was blocked to this session. Each claim below is marked with which it is.
Five names, each wrapping the last Each one exists because the one inside it ran out of road.
- 5names in the stack, from prompt engineering in 2022 to graph engineering in late 2026
- 2sources this session could actually read at source; the rest were blocked by the proxy
- ~100experiments autoresearch says it runs overnight, on a fixed five minutes each
- 12design primitives a public list breaks a harness into
1 · around 2022
Prompt Engineering 提示工程
- It started from one observation: the same question asked differently gives wildly different answers.
- Done badly: it answers the wrong question, and the format falls apart.
- It was not replaced. It went from lead role to component, and is still used at every step.
2 · mid-2025
Context Engineering 情境工程
- Why: a real application cannot be asked in one sentence. The model has to have seen the right documents and the earlier conversation.
- Shopify's CEO Tobi Lütke and Andrej Karpathy both switched to the term on X, and it displaced prompt engineering.
- Done badly: it misses a key file, or drowns in noise.
3 · around early 2026
Harness Engineering 駕馭工程
- Why: given the right material, the model still will not actually go and finish the job.
- The article's one-liner is Agent = Model + Harness, credited to Databricks' engineering blog. That post was not read here.
- Done badly: the model exceeds its authority, edits what it should not, breaks the team's rules.
4 · around mid-2026
Loop Engineering 迴圈工程
- Why: even fully equipped, a person still had to watch it and issue instructions step by step.
- The example is Karpathy's autoresearch — one of the few things on this page read at source. Details below.
- Its limits are equally clear: one thing at a time, one error and you start over, and earlier memory gets squeezed out by later.
5 · around late 2026
Graph Engineering 圖結構工程
- The chart holds only three things: steps that each do one job, the handoffs between them, and one complete record of progress.
- It buys four things: steps with no order between them run at once, an error can be recovered on the spot, the task can be halted at any time, and a different model can pick it up.
- Done badly: it cannot parallelise, cannot pause, and cannot be audited afterwards.
How firmly each part stands Read at source, reported but unchecked here, unconfirmed, guesswork — and two things that could not be reached at all.
| How firm | What | On what |
|---|---|---|
| Read at source | The autoresearch loop: modify the code, train for five minutes, check whether it improved, keep or discard, repeat. A fixed five minutes each, about 12 an hour, about 100 while you sleep. | The project's own README, fetched at source on 2026-08-15 |
| Read at source | The agent may edit train.py only and is barred from prepare.py, which holds data preparation and the scoring utilities. That separation stops it making the test easier instead of the model better. | The same README |
| Read at source | Harness engineering defined as the scaffolding around an agent — context delivery, tool interfaces, planning artifacts, verification loops, memory systems, sandboxes. The list breaks it into 12 design primitives. | The public awesome-harness-engineering list, fetched at source |
| Reported, not checked here | In mid-2025 Lütke and Karpathy both switched to “context engineering” on X. | As BusinessNext reports it. X was blocked to this session; neither post was read. |
| Reported, not checked here | Prompt and context engineering both live inside harness engineering; Agent = Model + Harness. | The article credits Databricks' engineering blog. It is the only authority the piece offers for the nesting, and it was not read here. |
| Unconfirmed | autoresearch “found settings humans had not noticed”. | The article says so. The project's README says no such thing and gives no results at all. |
| Unconfirmed | “Prompt engineer” was briefly treated as a fashionable new job title. | No source given. Widely repeated, nothing to check. |
| Guesswork | Harness around early 2026, loop around mid-2026, graph around late 2026. | These three dates appear only in the graphic, not in the article. Each carries “約” — approximately — in the original, and none is sourced. |
| Could not be reached | The article's own web page. | bnext.com.tw refused this session's proxy. The full text was pasted in by djTratoh on 2026-08-15, and that paste is the only copy this project holds. |
| Could not be reached | Who Josh C. Simmons is. | The article's source line reads only “X, Josh C. Simmons” — no link, no date, no title. No search from here identified them. |
Timeline
Everything in this story that has a date There are five. Three of the five layers have no checkable date at all, which is itself the point.
- 2022ChatGPT arrives. Both documents treat this as the start of the whole sequence, and of prompt engineering.
- mid-2025Shopify's CEO Tobi Lütke and Andrej Karpathy switch to “context engineering” on X. The article says only “mid-year” and gives no closer date.
- 27 Jul 2026BusinessNext's editorial desk finishes the concept graphic, drawing the five as boxes inside boxes and giving the last three an era label each, every one prefixed “約”.
- 28 Jul 2026The article publishes, bylined 陳建鈞. Two lines at its foot disclose that its sources are X and Josh C. Simmons, and that the first draft was written by AI.
- 15 Aug 2026Both documents are pasted into this project. bnext.com.tw refuses this session, so the text arrived by paste rather than by fetch.
The Argument
What the article argues Not replacement but containment — and how much room the article leaves around that.
The article makes one claim: these five are not successive generations replacing one another but layers wrapping one another, each shoring up the limits of the last. Its example is Databricks' framing — prompt and context both live inside the harness, the harness being the whole system around the model, with prompt and context as parts of it; the loop then automates that system, and the graph organises many looping models. So prompt engineering still works; it simply went from lead role to component. Then the article brakes on its own claim. There is no industry consensus on which layer sits above or contains which: some put loop above harness, some leave harness out altogether. Its own words are “目前還沒有定論” — not yet settled. Worth adding: two lines at the foot disclose that its sources are “X, Josh C. Simmons”, and that the first draft was written by AI before the journalist compiled and edited it. A piece about supervising AI output whose own first draft was AI-written is not a gotcha, but a reader should have it.
- 0industry consensus on how the layers stack, by the article's own account
- 2sources named at the foot: one platform, and one person's name with no link
- AIwrote the first draft. Compiled and edited by the journalist, 陳建鈞
What Others Add
The five, term by term What the article says, against what could be added after reading a source at first hand.
| Term | What the article says | What reading a source at first hand adds |
|---|---|---|
| Prompt | The instruction governing a single answer. | Nothing. No first-hand definition was reachable for this layer, and the article's account is uncontroversial. |
| Context | What the model can see and remember right now. Karpathy's definition, as rendered: putting exactly what the next step needs into what the model can see at once. | Nothing — and a caution: this is a Chinese rendering of an English post nobody here could read. X was blocked to this session. |
| Harness | Guardrails, tool-calling interfaces and a monitoring environment around the model. Agent = Model + Harness. | A public list breaks it down further, and says why: “Every component here exists because the model can't do it alone.” The twelve are below. |
| Loop | An automatic observe–act–check–retry cycle the model runs itself, as in Karpathy's autoresearch. | The README's own sentence: modify the code, train for five minutes, check whether it improved, keep or discard, repeat. Plus one separation the article omits and which matters most — below. |
| Graph | Organising several agents or loops into a directed graph or state machine. The chart holds three things: steps, handoffs, a record of progress. | Nothing. This is the article's most concrete section, and it comes with three working principles — below. |
How to draw the work as a graph — three principles The article presents these as a synthesis of developer-community practice, and sources them no further.
- 1One step, one jobEach step should do one thing, be testable on its own, and be replaceable whole without disturbing the others. A step with five things in it is just the old way again.
- 2Put the judgement at the handoffTwo kinds: a hard rule (“tests pass, deploy”) and an AI judgement (“does this complaint go to billing or to abuse?”). The developer's job is deciding which is which — and the points that need AI judgement are exactly where it will most often go wrong later.
- 3Record progress, and cap the spendAfter every step, write progress and output into a separate, fixed-format record. A failure then re-runs only the broken step, and the task can sit three days waiting for a person to approve something without stalling. If you cannot stop the model when cost hits its ceiling, it is not a system that runs itself — it is just burning money.
Two details the article does not carry, and both are worth having. Harness engineering is broken by a public list into twelve design primitives: agent loop, planning and task decomposition, context delivery and compaction, tool design, skills and MCP, permissions and authorisation, memory and state, task runners and orchestration, verification and CI integration, observability and tracing, debugging and developer experience, and human-in-the-loop. The list opens by saying why: every component exists because the model cannot do it alone. Note what that means: the agent loop is one primitive inside the harness — while in BusinessNext's arrangement the loop is the layer wrapped around the harness. That is precisely the disagreement the article calls unsettled. On loop engineering, autoresearch's most important design decision goes unmentioned: the agent may edit train.py and is barred from prepare.py, which holds data preparation and the scoring utilities. That separation blocks one very specific failure — an agent that can edit the exam will make the exam easier rather than the model better. Training runs to a fixed five-minute budget, scored on val_bpb, at a stated rate of about 12 experiments an hour and about 100 overnight.
- 12design primitives in a harness — one of which is the agent loop itself
- 1file the autoresearch agent may edit: train.py, and nothing else
- 5 minfixed training budget per experiment, giving about 12 an hour
Conclusion
What to take away Three things, and the first is about the picture.
The graphic is surer than the article, and the article is the one to trust
The graphic draws the five as a firm hierarchy and dates the last three. The article carries neither the dates nor the firmness, and says outright that which layer contains which is not settled. A public harness list puts the agent loop inside the harness — the reverse of the graphic, which wraps loop around harness. Read the picture as the answer at a glance and you miss what the article it illustrates actually says.
Three of the era labels are hedged, and appear only in the graphic
Around early 2026, around mid-2026, around late 2026 — each carries “約” in the original, none is sourced, and the article mentions none of them. Only two dates here are checkable: ChatGPT in 2022, and Lütke and Karpathy changing the word on X in mid-2025. The other three layers have no checkable date at all.
The most solid part is the three principles
Set the naming argument aside and the third principle is a test you can apply tomorrow: if you cannot stop the model when cost hits its ceiling, it is not a system that runs itself — it is just burning money. That holds without any of the five names, and without the industry agreeing on their order. So do the other two: one job per step, and the judgement at the handoff — where the judgement is, is where the errors will be.