The five rungs of agent engineering
Every team I've worked with that is "doing AI" is standing on one of five rungs. Knowing which one you're on tells you what to fix next — and, more usefully, tells you when tweaking the thing in front of you has stopped paying.
Rung 1 — Prompt engineering
You change the words. Add "think step by step", add examples, threaten the model, apologise to it.
This works, briefly. It is also where the overwhelming majority of organisational AI effort is still spent, long after the returns have flattened. The tell that you've outgrown it: your prompt is now longer than the task description, and each fix breaks a case you fixed last week.
Rung 2 — Context engineering
You stop editing the instructions and start editing what the model can see. Retrieval, tool results, schemas, prior state, examples selected per-request rather than pasted in.
This is a real step up, because it moves the problem from persuasion to information. Most "the model is dumb" complaints are actually "the model was never shown the thing." But you're still improving a single turn.
Rung 3 — Loop engineering
You let the system try, check, and retry. Validate the output, feed failures back, iterate to a stopping condition.
The unlock here is that correctness becomes a property of the loop, not of any single generation. A model that's right 70% of the time, inside a loop with a real verifier, is a system that's right far more often than that. The catch: your verifier is now the most important component you own, and a weak verifier makes the loop confidently converge on nonsense.
Rung 4 — Graph engineering
You stop treating the task as one loop and start treating it as a topology: stages, fan-out, barriers, specialised nodes, adversarial checks before anything commits.
This is where most genuinely hard problems get solved, because you can put verification where it belongs rather than hoping one agent remembers to be careful. It's also where you stop debugging wording and start debugging structure — "this finding was never independently checked" is a topology bug, not a prompt bug.
Rung 5 — Agentic ownership
The previous four rungs all make a single run better. The fifth removes you from the run.
The question stops being "how good is the output?" and becomes "who owns this task?" If the answer is "a person, when they remember to kick it off," that's the gap — no amount of rung 1–4 polish fixes a workflow that only executes when someone has the attention to start it.
Ownership means the work happens on a schedule or a trigger, runs to completion without supervision, and surfaces to a human only at checkpoints — the outward-facing, irreversible boundary. Sends, posts, purchases, deletions. Everything before that line runs unattended.
Why the order matters
The ladder isn't a maturity model to march through and graduate from. Rungs 1–4 still operate inside every rung-5 agent — an autonomous workflow still needs good context, a real verifier, and a sane topology. What the ladder gives you is a diagnosis:
- Output is wrong in a consistent way → rung 1 or 2. It's missing information.
- Output is wrong intermittently → rung 3. You need a verifier.
- Output is wrong in ways nobody caught → rung 4. Your topology has no adversarial node.
- Output is fine but nothing happens unless someone starts it → rung 5. It has no owner.
That last failure is the most common and the least discussed, because it doesn't look like an AI problem at all. It looks like a busy person with good intentions.
The uncomfortable implication
Rung 5 is mostly not a modelling problem. It's schedulers, queues, idempotency, error handling, notification design, and knowing which actions must never happen unattended. It is, in other words, infrastructure work — which is why teams with strong platform engineering tend to get further with agents than teams with strong prompt libraries.