Three years ago, "prompt engineer" was a real job with a real salary attached. It was relevant for about 18 months, the time it took the models to get good at asking for what they were missing. Then it became "context engineering." This year it turned into "harness engineering," and by summer people were saying "loop engineering."
Every few months the vocabulary resets, and every reset costs you a meeting where somebody asks whether you need one of these people now.
These are four stages of one idea, each moving further from the person typing the instructions: what you say, what the model sees, what surrounds it, what runs it. Knowing your stage tells you what to fix next, and it is rarely what your team thinks.
What each term means, and where most teams are
Prompt engineering (2022 to 2023) is about what you say. You are writing a better instruction. "Summarize this" gets you nothing useful, and a detailed request with a role, a format and a constraint gets you something you can send. It still works, and it is now a basic skill rather than a job.
Context engineering (2024 to 2025) is about what the model can see. Most bad output is not a phrasing problem. The model simply did not have your pricing sheet, your last quarter's numbers, or your house style. So the work moved to supplying the right information at the right moment: connecting the model to your documents, your systems, your data. Most companies I work with are somewhere in the middle of this stage right now.
Harness engineering (2026) is about everything around the model. The tools it can reach, the checks it has to pass, what happens when it fails, what it is allowed to touch. The working habit behind it: when the AI makes a mistake, you fix the environment so it cannot make that mistake again.
Loop engineering (2026) is about the system that runs the whole thing without you. Once the agent works, you stop typing at it. You build something that gives it work, checks the result, and sends it round again until it hits a stop condition you defined. Boris Cherny, who built Claude Code, put it plainly: "I don't prompt Claude anymore. I have loops that are running". It is a self feeding system designed around AI model.
The word "loop" is doing double duty here, and I have watched it derail a board conversation. The loop in "loop engineering" is the agent's own working cycle: try, check, adjust, repeat. It has nothing to do with "human in the loop," which is a separate governance decision about whether a person approves an action before it happens. You can run a well-designed loop with sign-off required on anything risky, and for most companies that is exactly the right setup. Cherny is describing where he sits, at the edge of the cycle rather than inside every turn, and he is not describing an absence of supervision.
Each new stage moves further from the model and closer to your operation. The skill being paid for keeps shifting from writing to designing.
Same model, 25 places up the leaderboard
In March 2026, LangChain took their coding agent from 52.8% to 66.5% on a standard industry benchmark for coding AI, moving from roughly 30th place to top 5. They changed nothing about the model. Every point of that improvement came from rebuilding the system around it: making the agent check its own work, telling it what environment it was operating in, and adding something to catch it when it got stuck repeating itself (LangChain).
Most leadership teams are still debating which model to standardize on. I believe that’s the wrong question. Your AI results are mostly a system-around-the-model problem, and that part is yours to build.
How this works in practice
A 10-person marketing team surveyed themselves before we started working together. Every person was using AI alone, as a chatbot, with paid licenses and the latest model. Outputs were inconsistent and off-brand, and the complaint that reached me was "the AI gives us generic work." After assessment, we figured what they needed was not better prompts but better context and control.
Context. One shared workspace holds who the company sells to, how it positions itself, and the reporting data with its targets. Every output starts from the same picture of the business, and the first drafts came back on-brand without anyone re-explaining the company.
Harness. The quarterly board report runs as a written routine with rules around it: it works on aggregated data only, because company policy says so; one person owns it; a person checks every number before it reaches the board. And when the first report came back unbranded, nobody rewrote a prompt. They added a rule so it cannot happen again.
The question I keep coming back to with clients: when the AI gets something wrong, what changes as a result? "The person tries a different prompt" means stage one, whatever the model costs. "We fixed the system so it cannot happen again" means you have moved outward, and that is the only move that compounds.
One cost warning for the last stage. Uber reportedly capped engineers at $1,500 a month for AI tooling after the annual budget went in four months. Loops that run themselves also spend by themselves.
Three moves for the next quarter
#1 Work out which stage you are honestly at. Ask a team using AI daily what happens when the output is wrong. Their answer places you on the ladder more accurately than any assessment questionnaire.
#2 Fix the layer you are actually on, not the one in the headlines. Most mid-market teams are stuck between stage one and stage two, and the highest-value work there is boring: giving the AI proper & safe access to how your business actually operates. Loop engineering is not your problem yet.
#3 Treat the model as the least interesting decision. It is the one thing you can swap later in an afternoon. The context, the checks, and the workflows around it are where the value accumulates, and they are the things nobody can sell you off the shelf.
Until next time,
Pooja
PS: If this was useful, please share it with a colleague whose team is still at stage one.