The New AI Job Titles, Decoded

Job titles are moving faster than the jobs underneath them. A posting for an "agent engineer" at one company describes work that another company calls applied AI engineering, a third calls forward deployed engineering, and a fourth just calls backend. Meanwhile some genuinely new specialisms have appeared that did not exist in any form three years ago.

For a job seeker this is not a semantic curiosity — it is a search problem. If you filter for the wrong words you never see roles you would be hired for, and you waste applications on roles that were never what the title implied. Here is the map.

Why the titles are such a mess

Three forces are colliding. Companies are inventing titles to signal that they are serious about AI, which inflates the vocabulary. Real new specialisms are emerging from actual technical need, which adds legitimate vocabulary. And recruiters are relabeling existing openings to attract a hotter applicant pool, which corrupts the vocabulary.

The result is that the title tells you roughly which neighborhood a job is in and almost nothing about the house. You have to read the requirements.

The titles worth understanding properly

AI engineer

The broadest of the new titles and the one most likely to appear on your search. The defining characteristic is that an AI engineer builds with models rather than building models: calling APIs, designing prompts and tool interfaces, wiring retrieval, handling failure modes, and shipping the surrounding product.

The term gained currency around 2023 as it became obvious that using a capable pretrained model is a different discipline from training one, and that the skills are mostly software engineering skills. This is the most accessible entry point for experienced engineers, because the hard parts are systems problems wearing new clothes. If you're weighing this against a research-adjacent path, we compared the two directly.

Forward deployed engineer

An engineer embedded with the customer, with the authority to change both the customer's implementation and the core product. Borrowed from Palantir, revived by AI companies because a general-purpose model needs substantial work to become a specific company's working system. It is the most customer-facing of the engineering titles, and the one with the widest gap between real versions and relabeled support roles — the full breakdown is here.

Evals engineer

The quietly essential one. Traditional software either passes a test or fails it. A model-backed system produces output that is better or worse, and "better" has to be defined, measured, and defended before anyone can safely ship a change.

Evals engineers build that measurement apparatus: test sets, scoring rubrics, human review pipelines, regression suites that catch quality drops a unit test would sail past. It is unglamorous and increasingly non-negotiable, because a team that cannot measure quality cannot improve it and cannot tell whether an upgrade helped. People arriving from QA, analytics, or research engineering are unusually well positioned and often do not realize it.

Context engineer

A newer term that emerged as the industry worked out that most model failures are not reasoning failures — they are failures to supply the right information. Context engineering is the discipline of deciding what goes into the model's working memory: which documents to retrieve, how to compress history, what tools to expose, how to structure the whole payload under a finite budget.

This is largely where the old "prompt engineer" work went, absorbed into something with more engineering substance. Retrieval, search, and data engineering backgrounds transfer well.

Agent engineer

Concerned with systems that take actions across multiple steps rather than returning a single answer. The hard problems here are recognizably distributed-systems problems: retries, partial failure, state across long-running tasks, permissions, and the question of what happens when step four of nine goes wrong in a way nobody anticipated.

If you have built orchestration or workflow systems, you already know most of this. The unfamiliar part is that one component is non-deterministic.

The title that faded

Standalone "prompt engineer" postings largely disappeared, and it is instructive to understand why. The role was real for a moment: models were brittle, phrasing mattered enormously, and knowing the tricks had genuine value.

Then models got substantially better at interpreting ordinary instructions, the tricks got written down, and it became clear that prompting was a skill within a job rather than a job. The work did not vanish — it grew up into context engineering and evals, which carry real engineering weight.

The lesson generalizes: titles built around a temporary technical difficulty tend to dissolve when the difficulty does. Titles built around a durable structural need — measurement, deployment, integration — tend to stick.

How to actually use this when searching

The practical advice is to search by the work rather than the title. Most job boards match on title strings, which means a single saved search misses most of what you want.

• Run several title searches in parallel, not one. AI engineer, machine learning engineer, forward deployed engineer, and AI product manager surface substantially different result sets at different companies.
• Search on requirement language too — retrieval, evaluation, agents, inference, fine-tuning — since these appear in the body of postings whose titles are unhelpful.
• Read the first three requirements before the title. Companies are careless about titles and specific about requirements, because the requirements were written by the person who needs the hire.
• Do not filter yourself out over the model layer. A large share of these roles want strong software engineers who are comfortable with non-determinism, not researchers.

Expect this vocabulary to keep churning for a while yet. Some of today's titles will be standard in five years and others will read like "webmaster" does now. The stable strategy is to stop optimizing for the label and get specific about the work you want to be doing — then search widely enough that a company's naming choices cannot hide the job from you.