One of the strongest AI improvements this year did not come from training a new model.

OpenAI recently ran GPT-5.6 Sol on a long-horizon evaluation. With its standard harness, the model scored 13.3%. After adding retained reasoning and better context compaction, the same model reached 38.3%.

Same model. Same task.

A different system around it.

That result explains where serious AI work is moving.

A powerful model inside a weak setup still forgets decisions, repeats completed work, calls the wrong tool, approves its own mistakes and stops because itsoundsfinished.

Put that same model inside a good harness, give it a real verifier, keep it running through a bounded loop and connect its work to other specialists through a graph and it becomes something very different.

This is why AI engineers appear to have abandoned prompting.

They have not stopped using prompts.

They have stopped sitting in front of the model and manually writing every next one.

The leverage has moved from telling AI what to do next to building the system that decides what happens next.

Five new terms, one growing system

The AI world keeps giving each new layer another name. It can sound like the previous skill has suddenly become useless.

That is not what happened.

Each layer simply controls a larger unit of work.

Prompt engineering is the message.It controls one request and one response.

Context engineering is the memory.It controls what the model can see, what should be retrieved, what should be compressed and what should be removed.

Harness engineering is the machine.It surrounds the model with tools, files, permissions, memory, logs, sandboxes, checks and budgets.

Loop engineering is the run.It keeps the agent acting, observing and correcting until a real completion condition is reached.

Graph engineering is the coordination.It decides which agents, loops, tools, reviewers and humans run next, what can run together and where failed work should return.

These layers do not replace one another.

Prompts live inside context. The harness assembles that context and controls the tools. The loop runs the harness repeatedly. The graph coordinates several loops and decision paths.

Zoom in and you find a prompt.

Zoom out and you see an entire AI operation.

Inside the full guide, you’ll learn where AI failures actually begin, how to build a safe harness, create a verified loop, turn it into a worker–reviewer graph, add real-world checks, control cost and permissions, and set everything up with practical prompts, commands, Claude Code, OpenAI Agents SDK and LangGraph.

Opinion AI has crossed100,000 subscribers in less than six months, and we are incredibly grateful for your support. To celebrate this milestone, our annual subscription is40% off for the next 24 hours, only $54 instead of $90. Join now and get full access to all our practical, step-by-step AI guides.

Get 40% off for 1 year

",{"_244":190,"_245":273,"_246":274},"7b928561-69ac-4554-b257-41a7164cf408",{"_19":273,"_31":275,"_192":276,"_33":277,"_215":278,"_251":278,"_21":222,"_252":279,"_254":280,"_208":40,"_68":-5,"_70":40,"_170":281,"_244":190,"_270":294},"TheSequence Opinion #904: The Age of Research Is Overrated. AI Engineering Is Winning","Ilya Sutskever's periodization of AI from research to scaling and back to research with bigger computers is appealing. However, frontier models still rely on familiar architectures like Transformers, with improvements stemming from data, context, reinforcement learning, and system integration. This suggests a shift towards engineering and industrial-scale application of existing research, where the surrounding system increasingly determines progress.","Why AI’s next breakthroughs may come from the learning loop around the Transformer—not from replacing it.","https://substackcdn.com/image/fetch/$s_!MfJS!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F917c6c6a-bf3b-49c5-aed8-66ba40e8a6a3_1672x941.png","2026-07-30T11:01:44","https://thesequence.substack.com/p/thesequence-opinion-904-the-age-of",{"_19":282,"_173":283,"_175":115,"_68":284,"_201":291,"_202":292,"_204":293},"dc9435f3a862eda370791cebacbc2cd1dda9ed75a20333cdc871175945952107","a9ed3df0ced363a33aa4226a94061292210cbef186e8c8f1e3ab6138e3a4e4f19c59c594074168119d69acaf3632db317e668b3a8a58982a420f33e7258c6cca",[285,286,287,289,290],[208,275],[35,278],[263,288,40,265],"30023:e461a563ccda0abe523218e4ea1da5a706532fe681a8b4d0feb5e0066a657583:7b928561-69ac-4554-b257-41a7164cf408",[198,40],[200,291,40],"e461a563ccda0abe523218e4ea1da5a706532fe681a8b4d0feb5e0066a657583","TheSequence Opinion #904: The Age of Research Is Overrated. AI Engineering Is Winning\nWhy AI’s next breakthroughs may come from the learning loop around the Transformer—not from replacing it.\n\n\nnostr:naddr1qvzqqqr4gupzperp543ueks2hefryx8yagw6tfcx2vh7dqdgkng0ad0qqe4x2avrqqjrwc3exgur2d3395mrjctr956r2df5943rydfh956rzcfhxymrgcmxxscrs3sjc3l ",1785409304,"

Ilya Sutskever recently offered a compact history of modern AI. From roughly 2012 to 2020, he argued, the field lived in an age of research. From 2020 to 2025, it entered an age of scaling. Now we are going back to research—only this time “with big computers.”

It is an appealing periodization. It also creates an immediate puzzle.

Open the release notes for almost any frontier model in 2026 and the architecture diagram looks strangely familiar. There is still a Transformer somewhere in the machine, often routed through a Mixture-of-Experts. The headline improvements are usually elsewhere: better data, longer context, stronger reinforcement learning, synthetic tasks, tool use, memory, verification, adaptive reasoning budgets, and agent orchestration.

This does not look like the arrival of a new neural species. It looks like Formula 1. The car still has four wheels and an engine. Yet enormous gains come from aerodynamics, energy recovery, tire chemistry, telemetry, software, and pit strategy. The chassis matters. The system around it increasingly decides the race.

So which era are we in: research or engineering?

Probably both. The age of research has returned, but much of that research is now expressed as industrial-scale engineering.

“Scaling did not end. It escaped.”

The Recipe That Ate the Field

",{"_244":190,"_245":296,"_246":297},"573e96d8-b1cd-45f0-9868-64955019417d",{"_19":296,"_31":298,"_192":299,"_33":300,"_215":301,"_251":301,"_21":222,"_252":302,"_254":303,"_208":40,"_68":-5,"_70":40,"_170":304,"_244":190,"_270":316},"Graph Engineering: The Next Step After Prompts and Loops","Graph engineering is presented as the evolution beyond prompt and loop engineering for AI, focusing on the coordination of multiple agents to manage complex tasks. It introduces a system where nodes perform work, edges dictate workflow, and state management ensures continuity, moving towards autonomous operations.","Prompt engineering told AI what to do. Loop engineering kept it working. Graph engineering connects the whole operation.","https://substackcdn.com/image/fetch/$s_!Y56s!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F32a658b5-1a9c-48f5-a866-2ccac44e4907_1672x941.png","2026-07-21T00:49:01","https://emergingai.substack.com/p/graph-engineering-the-next-step-after",{"_19":305,"_173":306,"_175":115,"_68":307,"_201":26,"_202":314,"_204":315},"f2272fbdc8e896f65bc37bbd85d45203d901c16deabde3fa200fbcaf38cc3d7c","a4563c1f6b86abdab3a6df0f669d6e20bd8312dd0d02bc9328a4f2e8b1b5bcaccf809e4eda7bface7d72fbb62ad89b8b3d98768f4fa7ec6ad40ab8d548329b30",[308,309,310,312,313],[208,298],[35,301],[263,311,40,265],"30023:983dd797fe72ad62fc8e3ece7eed1f126550972885c8f835d9af5528dad655a3:573e96d8-b1cd-45f0-9868-64955019417d",[198,40],[200,26,40],"Graph Engineering: The Next Step After Prompts and Loops\nPrompt engineering told AI what to do. Loop engineering kept it working. Graph engineering connects the whole operation.\n\n\nnostr:naddr1qvzqqqr4gupzpxpa67tluu4dvt7gu0kw0mk37yn92ztj3pwglq6ant649rddv4drqqjr2denv5unvepc943rzcmy956r2e3s95unsd3c95mrgwf4x5crzwf5xymkgku4uyn ",1784594941,"

Prompt engineering told the model what to do. Loop engineering taught it to keep working. Graph engineering decides who works, who checks, what happens next, and where a human takes control.

On July 17, Peter Steinberger posted one question:

It travelled because builders recognized the problem immediately. We had finally learned how to keep one agent working without writing the next prompt ourselves. Then we discovered that ten good loops still leave one person carrying results between them, deciding what starts next, checking which output can be trusted and restarting whatever failed.

The work became autonomous. The coordination did not.

I had already run into this in my own setup. For serious tasks, I stopped using one agent and began using three.

The first is theworker. It researches, writes, builds or fixes.

The second is theinvigilator. It does not help the worker. It sees the output with fresh eyes and looks for weak reasoning, missing evidence, shortcuts and fake completion.

The third is theevaluator. It checks both of them. Did the worker complete the task? Did the invigilator catch the important problems? Is the result ready, or should it return for another round?

That is the practical jump from a loop to a graph.

A prompt produces an answer.
A loop pursues a pass condition.
A graph runs the operation

The loop was not wrong. It was promoted.

A loop remains one of the best structures in AI. Give an agent a goal, let it act, test the result, return the failure, and stop when the test passes or the budget ends.

Anthropic’s evaluator–optimizer pattern uses exactly this shape: one model generates while another evaluates and feeds back corrections.

The weakness appears when the work contains several goals.

A launch may need research, copy, code, legal review and approval. A code change may need architecture analysis, implementation, security review, tests and release. A customer request may need classification, policy lookup, account data, a decision and a human sign-off.

You can build a loop for each part. But somebody still has to connect them.

Graph engineering turns that hidden coordination into a visible system:

Nodesdo work. A node can be an agent, script, API call, test, database query or human review.
Edgesdecide what happens next. They carry data, dependencies, permissions and failure routes.
Stateremembers what happened outside any single chat window.
Gatesdecide whether work may continue.
Cyclessend failed work back without restarting everything.
Terminal statessay done, rejected, expired or needs a human.

Inside the full guide:how Graph Engineering moves beyond prompts and loops, how to build worker–reviewer–evaluator systems, run parallel checks, route failures, add human approval, define node contracts, choose the right model for each job, build it in Claude Code or LangGraph, and stop the whole system from becoming confidently wrong.

",[],"selectedArticleImage","rootArticleImageUrl","consensusHtml","tweets",[],"userSchema","relatedArticles",[247,274,297],"actionData","errors"]

AI Development Focus Shifts From Prompting to 'Graph Engineering'

Story coverage