Andrew Ng dropped a number this week: 90% of enterprise Agent projects stall at deployment. But a recent long-form post on Juejin, "From Prompt to Graph," gave us another angle — it's not that the technology falls short; it's that the engineering language itself is still shifting gears fast. In that span, five layers of terminology have taken turns on stage — Prompt, Context, Harness, Loop, Graph — and none has knocked the previous one down; each has just pushed the leverage one notch higher.

What This Is

The article strings together the five layers of AI engineering along a timeline:

Prompt (2020–2023): Learning to "speak to the model." Few-shot and Chain-of-Thought (letting the model write out its reasoning steps before answering) turned Prompting from a casual question in a chat box into real engineering work.

Context (2024–2025): What you stuff into the window matters more than how you ask. In late 2024, Anthropic open-sourced MCP (Model Context Protocol, a universal interface protocol for models to call external tools), unifying the way tools, databases, and repositories are connected.

Harness (early 2025–2026): The entire "racetrack" wrapped around the model — tool permissions, sandboxes, hooks, skill packs, session resets, evaluation mechanisms. Anthropic's engineering blog puts it bluntly: every component in the Harness patches a gap the model still can't fill on its own.

Loop (June 2026): Stop personally acting as the Prompt operator. Boris Cherny, head of Claude Code, has publicly said he no longer writes Prompts himself — he writes Loops, using scripts to cycle through Prompting an Agent. Google engineer Addy Osmani distilled this practice into the Loop Engineering framework.

Graph (July 2026): How multiple Loops get wired together. Who goes first, who goes after, how to branch, how state gets passed, where failures retreat. Tools like LangGraph have existed for a while; what's new is that people are starting to discuss "whether multi-Agent relationships should be drawn out clearly" using the same vocabulary.

Industry View

The pro side's judgment is clear: as models get stronger, the old scaffolding should come down — but the parts models still can't handle (long-horizon task management, context filtering, product acceptance) all rely on Harness, Loop, and Graph to fill in. The Anthropic engineering team's line gets quoted repeatedly: "Every component in the Harness patches a gap the model still can't fill on its own."

But the dissent is equally sharp. The article itself points out that most Coding Loops' "done condition" still stops at lint passing (code style checks passing) and unit tests green (unit tests passing). Anyone who has shipped an App knows: code that can merge doesn't mean the product has passed review — clicking through the UI, running it on a real device, still often requires a human. On the surface, Loops are spinning; the product-validation branch is often broken. In other words, the engineering language has been upgraded five layers, but the question of "who signs off" has never been properly answered.

An even more hidden risk: the speed of terminology iteration itself. Five layers in two years means a team betting on Harness today may have to rewrite their architecture six months from now. For resource-constrained enterprise IT departments, that's a non-trivial gamble.

Impact on Regular People

For enterprise IT: Don't get pulled along by new buzzwords. Ask first: "Is our core pain point that Prompts aren't written well enough, that Context is being stuffed in wrong, or that we're short on people for the acceptance step?" Then decide which layer to invest in.

For individual careers: "Knowing how to write Prompts" is no longer a scarce skill. What's scarce is knowing "when a human should take over" and "how to validate AI output" — precisely what Harness and Loop address.

For the consumer market: In the short term, the AI tools you use won't suddenly get better because the terminology has iterated. The perceptible changes will come slower: customer service bots that finally stop giving irrelevant answers, office software that can run an entire workflow end-to-end — but only if enterprise IT is willing to pay for the underlying scaffolding.