We index around 1.1 million UK careers pages, and exactly three employers currently put “agentic” in a job title. Plenty of British teams build these systems under the title AI engineer, but the word has not settled — so a search built on it returns three adverts and a lot of consultants.
The word also covers both a chain that calls two tools in a fixed order and a system that runs unsupervised for an hour deciding what to do next. Those are different problems, and somebody excellent at the first can be a liability on the second.
Pin down how much autonomy you are actually buying
Before you write the advert, answer this: how many decisions does the system make that you did not enumerate in advance?
If the answer is none — the model classifies, a tool is called, the output is formatted — you have a pipeline with a language model in it, and a competent backend engineer will build it. If it chooses which tools to call, how often, and when it has finished, you are buying a control loop whose state space you cannot write down, and the skills sit closer to distributed systems than to prompting.
How long it runs before a human sees anything
Three seconds and one turn is forgiving. Twenty minutes over forty steps is not, because by then the system has already acted several times. Ask for the longest-running agent they have operated and what proportion of runs failed. Anyone who has done it has a figure.
A tool call is a network call, and it will fail mid-plan
The interesting failures are not hallucinations. They are the ordinary indignities of calling other people’s software: a timeout, a tool returning success with an empty body, an argument the model invented because the description was ambiguous.
Listen for how the loop behaves when a middle step comes back wrong. The honest answers are specific: retry with the error fed back into context, retry with a narrower instruction, fall back to another tool, or escalate to a person. The answer to worry about details the happy path and treats errors as the framework’s problem.
Then push on recovery. If a step three actions deep produced a plausible but wrong result, does the agent notice at step seven, and what happens to the four actions since? People who have supported one talk about checkpointing state so a run resumes rather than restarts.
Idempotency, because an agent that writes will write twice
Here is the question that separates people who have put an agent in front of real systems from people who built demos: what happens when yours sends the same email twice?
Every action that writes — a refund, a ticket, a calendar invitation, a payment — will eventually be attempted twice, because the model retried, or the run resumed, or a timeout hid a successful call. The engineers you want solved this the boring way: a key derived from the intent rather than generated fresh, a log of completed actions checked before acting, a tool layer that refuses duplicate side effects.
Ask which actions they left reversible, and why. Making an agent ask permission before a destructive step is easy; knowing which steps qualify is the judgement you are paying for.
The runaway loop, and the budget that stops it
A misbehaving agent is not only wrong, it is expensive. A loop that never recognises it is done keeps calling the model with a context that grows every turn, so cost scales near the square of the step count. Teams routinely discover this from an invoice.
Ask what the ceiling per run was and what enforced it. Good answers involve hard caps on steps, tokens, wall-clock and spend, enforced in code, not requested politely in the prompt, plus an alert when a run hits one, because that is a bug report. Ask what the p95 cost per run was; a candidate who offers a distribution rather than an average has been paying attention.
Reading back a run that went sideways
A forty-step run leaves no account of itself unless somebody built one, which makes this the hardest part of the job to fake. Ask how they established what one bad run actually did.
You want to hear about a trace: each step, the tool, the arguments, the raw response, the model’s stated reasoning, tied to an identifier you can search from a support ticket. Then the harder half — what did they do with it? Replaying recorded runs against a changed prompt, a library of failure cases that became a regression suite. If your real difficulty is knowing whether the system is improving, a model evaluation engineer may be the better first hire.
Three adverts is not a benchmark, but here are the numbers anyway
Three employers using the word in a title is not a sample anyone can derive a reliable band from. What follows is inferred from adjacent markets and should be read as indicative.
For a mid-to-senior engineer who has genuinely run one in production, roughly £85,000 to £120,000 permanent, higher in London where the agent is the product. Contract, indicatively from about £600 a day, the top of that range where the work touches money or regulated actions. The premium over a general AI engineer brief is real but narrower than the noise suggests.
Because the title barely exists, do not search for it. The people you want appear on CVs as backend, platform or LLM engineers who shipped something that plans and acts. Describe the system in the advert, not the label.
When a state machine would have done the job
Now the question most likely to save you the hire. If you can draw the process on a whiteboard and the boxes never change between runs, you do not want an agent. You want a workflow with model calls at the steps that need judgement: cheaper, testable, and it will not surprise you at three in the morning.
The honest case for autonomy is a long tail you cannot enumerate — thousands of differently shaped requests, or an environment changing faster than you can maintain branches. The dishonest case is that it sounds more advanced than a queue and a switch statement.
Anyone worth hiring says this before you do. Put your real process in front of them and ask where they would remove it; the ones who remove it from two-thirds of the diagram have paid for the alternative.