Loop Engineering Is the New Senior Skill, and It Raises the Bar on Your Engineers

A clever prompt can make an agent look capable once. A well-designed loop makes it dependable every time. Xun Wang, chief technology officer at Bloomreach, a martech and commerce-experience company, argues that the best engineers are already focused on the second.

Wang wrote on LinkedIn that "the best engineers have stopped prompting their AI agents. That sounds backwards until you see what they do instead: they design the loop the agent runs inside. It is the shift from prompt engineering to loop engineering, and it quietly redefines the job." His claim isn't that prompting stopped mattering, it's that the hard part of the job moved somewhere else.

The loop behind the output

The loop is the system the agent operates inside. It's the context the agent gets handed before it starts, the boundaries on what it's allowed to touch, the checks that catch a wrong step before it ships, and the feedback that makes the next run better than the last. None of that is phrasing. It's architecture, the same design thinking that has always separated engineers who can build a system from engineers who can only write a function. Wang's point is that that work is becoming the center of the job instead of something you do around the edges.

The lazy version of this story runs the other way. If an agent writes the code, the reasoning goes, then you need fewer engineers and less seniority, because anyone can ask a model for a function and paste the result. That reading confuses the request with the result. A prompt is a request, and a loop is the structure that decides whether the request comes back as something you can trust, which is squarely a senior engineer's job to build. When Wang says the best engineers design the loop, he's naming the part of the work that got harder, not the part that went away.

'Almost right' still breaks in production

The reliability numbers are why the loop matters. In a randomized controlled trial from the research group METR, 16 experienced developers worked 246 real tasks in large open-source codebases they already knew well, and they took 19% longer to finish when they were allowed to use AI tools than when they weren't. The developers themselves guessed the tools had sped them up by about 20%. The distance between what the agent felt like and what it actually delivered is the distance a good loop exists to close, through the checks and the measurement a lone prompt never supplies.

Working engineers already feel where the cost lands. In Stack Overflow's 2025 survey of more than 49,000 developers, the biggest single frustration, named by 66%, was AI output that's "almost right, but not quite," and 46% said they don't trust the accuracy of what the tools produce, up from 31% a year earlier. The near-miss is the failure mode that matters, because it slips past a careless glance and breaks later in production. Catching it isn't about asking the model more politely. It's about designing evaluation and feedback into the loop so the almost-right answer gets caught before a person has to spot it by hand.

Better loops beat better prompts

The management problem begins where the prompt ends. As agents get handed to every team at once and everyone works the same models, the edge stops being who writes the cleverest prompt, since that skill levels out fast. The edge becomes who designs the better loop, the bounding, the checking, the context, and the feedback that turns a capable but unreliable agent into a dependable one. That's harder work than prompting, and it rewards exactly the experienced judgment a leader pays the most to hire. Instead of thinning out the case for strong engineers, the agentic shift raises the bar they're measured against.

That's the honest version of Wang's argument. The job didn't get smaller when the agent started doing the typing. The valuable part climbed, from the words you feed a model to the system you build around it, and that system is still an engineer's to design. A model can run the loop. It still takes a good engineer to draw it.