Blog | WeScale

AI Coding Agents: From Writing Prompts to Designing Systems

Written by Emmanuel Sciara | Oct 19, 2026, 9:57:20 PM
The engineering frenzy: one summer, two new words 
On June 7, 2026, five days after Boris Cherny made a statement about loops, Peter Steinberger posted on X:  

«  Here’s your monthly reminder that you shouldn’t be prompting coding agents anymore. You should be designing loops that prompt your agents.  »

That same day, Addy Osmani popularized it as "loop engineering", and by June 30 Anthropic had published a guide to getting started with loops.
On July 18, Steinberger asked the obvious follow-up: 

« Are we still talking loops or did we shift to graphs yet? »

In a single summer, a term broke through, took hold and was already being questioned.
Caught in the hype spiral, it is easy to believe that each new word makes the last one obsolete. The better reading is the opposite: from prompt to context to harness to loop, and maybe to graph, each new term adds a ring around the last one instead of replacing it.
Each term defines an engineering practice that opens new possibilities. The bottleneck moves outward, and the human steps further back each time. And every ring is still there.

Five rings
For each ring, one question matters more than the name: where does the human sit?
The dates in the headings below don't mark when each practice was born (some were in use well before), but when it took over from the previous one as the main technical bottleneck.
five nested rings around the model 
Prompt (2022–2024)
In the chat era the unit of work was the prompt. You typed a request, read the answer, typed the next one. Prompt engineering was the craft of phrasing: roles, examples, "think step by step". The bottleneck was the wording, and the human supplied it.
With prompt engineering, the human sits in the conversation, at every turn.

Context (mid-2025)
On April 15, 2025, Dex Horthy coined the term "context engineering" in "12-Factor Agents", his guide to building reliable LLM applications. Two months later it went mainstream: on June 19, Tobi Lütke said he preferred it to "prompt engineering", and Andrej Karpathy agreed on June 25.
Their point: models had become good at following instructions, so what limited them was what they could see. The right files, docs, history and tools in the window mattered more than clever phrasing.
With context engineering, the human sits beside the model, curating what goes in.

Harness (early 2026)
By then agents no longer answered one message at a time; they worked on their own for minutes, sometimes hours, before checking back. On February 5, 2026, Mitchell Hashimoto described his rule: when the agent makes a mistake, fix the environment so it cannot happen again. Six days later, OpenAI's Ryan Lopopolo described a product of about a million lines with zero hand-written code, built in five months by a team that grew from three engineers to seven. Their work was tests, linters, docs and conventions: the agent's environment.
With harness engineering, the human sits outside the session, shaping the agent's working environment and its tools.

Loop (June 2026)
 A harness makes one session reliable. Someone still has to decide what runs next, check the result and start again. Increasingly, that someone is code. Boris Cherny put it bluntly in his interview with WorkOS on June 2:

«  I don't prompt Claude anymore. I have loops that are running. They're the ones that are prompting Claude and kind of figuring out what to do. My job is to write loops. — Boris Cherny (11:42) »

A loop is the outer cycle that re-prompts the agent, verifies its work and re-runs it.
Stepping back has a price, though. The further the human sits from the code, the more the agent's "done" becomes a claim rather than a proof, and the less the team understands what it ships.
In loop engineering, the human sits on the loop, not in it, designing triggers, checks and stop conditions.

Graph? (July 2026, contested)
One loop handles one stream of work. Run ten and something has to coordinate them: who works on what, what state they share, where a human signs off. That is what Steinberger's July 18 question pointed at. Whether "graph engineering" survives as a term is open; skeptics note that agent frameworks have done this for years.
If graph engineering is the next ring, the human designs the organization of loops.

Nesting, not replacement
Read the terms as a succession of groundbreaking approaches you must adopt on the spot, and each one seems to erase the last.
Read them as rings and everything adds up. Even the words are older than their fame: each was coined days, months or even years before someone famous made it trend. (read this article: Who Coined What: The Origins of the Agent-Engineering Terms)
In practice:
  • A graph is still built on loops.
  • Each loop runs an agent inside a harness.
  • Much of a harness's job is to feed the agent the right context.
  • And that context still holds prompts: the task description, the AGENTS.md file, the instruction the loop sends on every turn.
Prompts did not disappear when context engineering arrived. They became a subset of context. The same thing has happened at every step since. So the older skills do not expire. A badly worded prompt inside a well-designed loop is still a badly worded prompt, now repeated a hundred times overnight.
What changes is where the leverage sits. Each ring is where the bottleneck moved once the inner one was good enough.

Find your ring
The useful question is not which term is current. It is which ring your bottleneck lives in, so you can work there:
  • If the agent misreads what you ask, work on the prompt.
  • If it lacks the facts, work on context.
  • If it makes the same mistake twice, work on the harness.
  • If you spend your day deciding what runs next, work on the loop.
  • If you are juggling five loops by hand, graphs may be your answer.
The next three posts take one ring each: [the harness], [the loop], and [what might come after it].
The next term will arrive soon, probably before this series ends. When it does, don't ask what it's called. Ask where it puts the human.
Sources
View full post