What jobs can you get if you know agentic AI and software development?

You can write software. You've built three or four agents that actually do something useful, wired up a few tools, wrestled with retries and token costs. Then you open a job board, search for something that matches, and get a mess of titles: AI engineer, agentic AI engineer, forward deployed engineer, AI solutions architect, automation specialist, machine learning engineer (which is usually not the same job at all).

So the question I keep getting from developers is fair: what job can you actually get if you know agentic AI and you're good at software? I hire and work alongside people in these roles, and I've been on the buying side of them too, so here's what I see in the market rather than what the course sellers say.

Is there real demand for agentic AI skills, or is it just noise?

Both, honestly. The demand is real but narrower than the headlines suggest.

Indeed's Hiring Lab tracks the share of US job postings that mention AI, and it hit a high of 4.2% at the end of 2025 while overall hiring stayed weak. In other words, AI-mentioning roles are growing against a flat or falling market. Indeed also found that US software development postings grew by almost 15% in the period after Claude Code launched in early 2025, while total postings fell by 7%, which is the opposite of the "AI kills developer jobs" story people expected.

On the buyer side, one 2026 hiring survey citing McKinsey's State of AI research put the numbers at roughly 23% of organisations scaling an agentic system somewhere in the business and another 39% experimenting. That gap between experimenting and scaling is where your job lives. Most companies can get a demo working. Very few can get one into production and keep it there.

One caveat before you plan a move: Lightcast's analysis for the Stanford AI Index shows US AI hiring is still concentrated in a small number of hubs, with California alone accounting for around 171,000 AI job postings in 2025. If you're not in a hub, remote and contract work matters more for you than for the average developer.

What job titles actually exist for agentic AI and software skills?

Here are the roles I actually see people get hired into. The titles vary by company. The work doesn't vary as much as the titles suggest.

1. AI engineer or agentic AI engineer

The default title. You build systems on top of models rather than training them: tool calling, retrieval, orchestration, evals, guardrails, cost and latency control. Strong general software skills matter more here than machine learning theory. If you can design an API, reason about queues and idempotency, and debug a distributed system, you're most of the way there.

2. Forward deployed engineer or AI solutions engineer

This one has exploded and it's underrated. You sit with a customer, learn their messy process, and build the agent inside their environment. Half engineering, half consulting. AI startups hire hard for it because their product only works when it's shaped around a specific workflow. If you like talking to people and hate writing specs nobody reads, this is the best-paid version of that personality.

3. AI platform or infrastructure engineer

Once a company has ten agents running, someone has to own the boring layer: model routing, spend limits, logging, evaluation pipelines, secrets, access control, an MCP server or an internal tool registry. This is a proper software engineering job with AI attached, and it tends to be the most durable of the lot.

4. AI product engineer at a startup

Small team, you own features end to end, the agent is the product. Expect wide scope and less structure. Good if you want to learn fast, risky if the company's only moat is a prompt.

5. Internal AI lead at a non-tech company

Insurance brokers, clinics, logistics firms, law firms, universities. They have a backlog of manual work and no one who can build. These roles are less glamorous and often less well paid than a frontier lab job, but there's far less competition and you'll ship things that matter within weeks. If you already work at one of these companies, the fastest route into agentic AI work is not a new employer, it's a project.

6. Consultant, freelancer or small studio

This is what I do. You sell outcomes rather than hours: an agent that handles inbound quotes, an internal ops assistant, an MVP for a founder who needs to raise. It pays well when you can scope tightly and say no to bad projects, and it pays badly when you can't. If you're weighing this up against a salary, my post on in-house vs freelance vs agency covers how clients think about the trade-off, which is useful intelligence when you're the one being hired.

What do these agentic AI jobs pay?

Pay is wide and location-dependent, so treat any single figure with suspicion. As a rough anchor, Glassdoor's US data puts the typical AI engineer range between about $114,000 and $181,000 a year, with an average around $143,000. London, Toronto, Sydney and Dubai run lower in raw numbers, though Dubai's tax position changes the comparison. Frontier labs and well-funded startups pay far above any of this, and they are also the hardest roles to get.

If you go independent, price by project rather than by hour. For context, my own starting points are $1,500 for a short consulting engagement, $4,000 for an agentic system build over two to six weeks, and $6,000 for an MVP sprint. Those are my numbers, not a market rate card, but they'll tell you roughly what small and mid-sized businesses are willing to spend on a first real agent. Anyone promising you $30,000 retainers from cold DMs is selling a course.

What do employers actually test for?

Almost nobody senior is hiring for prompt skill any more. Here's what comes up in real interviews and real client conversations:

• Evals. How do you know the agent got better after your change? If your answer is "I tried a few examples", you're a hobbyist. A test set, a scoring method and a regression check is the bar.
• Failure handling. What happens when the model invents a customer ID, the third-party API times out, or the same webhook fires twice? I wrote more about this in can you trust AI agents.
• Cost and latency. Can you explain the per-run cost of your system and where it goes? Teams get burned by agents that work beautifully and cost $4 a conversation.
• Integration work. Most of an agent build is plumbing into CRMs, ticketing systems, databases and half-documented internal APIs. Boring, and it's where the value is.
• Data handling. Knowing where the data goes, what the provider retains and what your client is legally on the hook for. The rules differ across markets: GDPR in the UK and EU, a patchwork of state laws in the US, PIPEDA in Canada, the Privacy Act in Australia, the PDPL in the UAE. You don't need to be a lawyer, but a developer who can answer "where does this data actually sit" gets hired over one who can't.
• Judgement about what not to build. The most valuable thing I do on a project is talk someone out of an agent when a scheduled script would do. Saying that in an interview signals seniority faster than any framework name.

What's overhyped in the agentic AI job market?

A few things I'd tell a friend over coffee.

Certifications aren't doing much. I've never once chosen someone because of an agentic AI certificate, and I've never been asked for one. A public repo with a working system and a short write-up of what broke beats any badge.

Framework loyalty is a trap. Tools churn every few months. The transferable skills are system design, debugging and knowing when a model is the wrong tool. If your CV reads like a list of framework names with no outcomes attached, it looks thin.

"Prompt engineer" as a standalone job has mostly faded into normal engineering roles. Prompting is a skill inside a job, not a job.

And demo-ware is cheap now. Anyone can produce a slick agent video. What's scarce is someone who has kept a system alive through three months of real users, weird inputs and a model deprecation. That's also why so many projects fall over after launch rather than during the build, which I wrote about in most AI projects don't fail during the build.

How do you get hired without an AI job on your CV?

The route I'd take if I were starting again this month:

1. Pick one painful, boring process you understand well. Invoice chasing, support triage, lead qualification, report generation.
2. Build the agent that handles it properly, including error paths, logging, a human approval step and a cost ceiling. Ship it somewhere real, even if that's your current employer or a friend's business.
3. Measure something. Hours saved, response time, error rate before and after. Numbers turn a side project into evidence.
4. Write up what went wrong, not just what worked. Hiring managers read that and recognise someone who has actually done it.
5. Do two or three paid projects at a low price to build proof, then stop discounting.

That sequence works whether you want a salary, contracts or your own studio. It's also roughly the checklist founders use when they hire, which I set out in how to hire an AI developer. Reading the buyer's side of it is one of the cheapest edges you can get.

The short version

If you can build agents and you can ship software, you're in a decent position, but not because "AI" is on your CV. You're in a decent position because the market is full of companies stuck between a demo and production, and very few people can close that gap. Aim at that gap. Choose the role that fits your temperament, whether that's forward deployed work with customers, platform work behind the scenes, or independent projects. Then get something real into production and write about it honestly.

And if you're reading this from the other side, as a founder trying to work out which of these people you need and what to hand them first, get a build plan and we'll map it out together.