Choosing Your AI Specialisation: Agents, Evals, Infra or FDE

What you need to know

Two years ago, "AI engineer" was one job. As of July 2026, it is an umbrella over several distinct hiring tracks, each with its own interview rubric, its own pay band and its own definition of proof. LinkedIn still ranks AI engineer among the fastest-growing titles in the market — but underneath that headline, companies from Anthropic and Salesforce to EY, Deloitte and Accenture have spent the first half of 2026 posting job titles that did not exist two years ago. That split changes the career question. It is no longer "how do I get into AI?" — it is "which of these tracks am I actually running on, and can anyone tell from the outside?"

• The generic title has split into at least five hireable tracks: agentic engineering, evals, infrastructure, context engineering and forward-deployed engineering.
• Generalists still get hired — smaller teams genuinely need range — but the top of every band goes to people who picked a direction and built visible proof in it.
• Pay has diverged sharply. One US career guide reports evals-engineer total compensation of $230k–$650k+ at frontier labs, while generic AI-engineer medians in every market sit far below that.
• Your current role largely predicts your most natural track — backend leans agentic, DevOps leans infra, QA leans evals, domain experts lean context or forward-deployed.
• You can pivot in roughly 90 days without quitting, because every track's entry ticket is one substantial artefact plus a public writeup, not a qualification.

The tracks that actually exist in 2026

These five are not predictions — each is verifiable in live job postings and hiring surveys as of July 2026. The boundaries blur at smaller companies, but at any organisation large enough to run a structured hiring loop, these are now separate roles with separate rubrics.

Agentic AI Engineer

The person who takes a capable model and turns it into a production agent system: orchestration, tool design, state management, guardrails and the observability that makes any of it debuggable. Enterprise agent rollouts drove much of 2026's hiring surge — Forbes' June 2026 survey of Box, McKinsey and LinkedIn hiring data catalogued a wave of new agentic titles, and one 2026 tracker put agentic AI job postings up roughly 280% year on year. In India, the demand shows up strongly in Bengaluru and Hyderabad GCCs building agent platforms for global parents; in the UK, it clusters in London consultancies and fintechs wiring agents into regulated workflows. The day-to-day is less "clever prompting" and more distributed-systems thinking applied to a non-deterministic component.

AI Evals Engineer

The measurement discipline. Evals engineers own the systems that decide whether a model or agent change ships or does not — task-specific test suites, error analysis, regression detection and the judgement calls about what "good" means for a given product. It is the track with the most dramatic reported pay at the top: one 2026 career guide aggregating employee-reported US figures puts total compensation at roughly $230k–$340k for mid-level, rising to $460k–$650k at staff level, with frontier labs paying above those ranges. Treat that as what it is: a single source, US-specific, and heavily equity-weighted — not a promise, and emphatically not an anchor for an offer conversation in Chennai or Manchester. What is verifiable across sources is the direction: because evals sit on the critical path for every model release, the role is scarce, senior-skewed and consistently priced at or above equivalent product engineering.

AI Infrastructure Engineer

Serving, throughput and cost. This track owns inference platforms — vLLM and its rivals, GPU scheduling, caching, routing and the unglamorous economics of tokens per second and cost per million tokens. Recruiting firms describe MLOps and AI-infra roles as among the hardest to fill of 2026, and the job specs read like a DevOps CV with a serving layer on top: Kubernetes operators, CUDA-adjacent tooling, automation in Python or Go. That makes it the most natural transition in this article — a DevOps or SRE engineer is perhaps one published benchmark away from being a credible candidate. Both markets need it: Indian GCCs are standing up in-country inference for data-residency reasons, and UK enterprises are doing the same under their own sovereignty and procurement pressures.

Context Engineer

The newest title on the list, and the one to hold most lightly. Context engineers design everything a model sees on each call — retrieved documents, tool definitions, memory, and above all curated, version-controlled context: the CLAUDE.md files, rules folders and knowledge bases that move agent reliability more than any prompt tweak. The title surfaced in late 2025; by early 2026 ODSC was cataloguing it among the year's emerging roles and Gartner had published a formal definition. Real listings exist, but many companies still fold the work into agentic roles — so build the skill and the proof, and treat the title itself as a bonus rather than a plan.

Forward Deployed Engineer

The customer-facing track. Palantir originated the FDE role around 2005 to embed engineers inside intelligence and defence customers, and reportedly ran more FDEs than conventional software engineers for its first decade. In 2026 the model has gone mainstream: OpenAI and Anthropic have both built out FDE teams, Google and EY are hiring the title, and Anthropic's own Applied AI postings describe FDEs embedding directly with strategic customers. An FDE ships working systems inside someone else's organisation — half consultant, half engineer, measured on customer outcomes rather than internal velocity. If the shape of the role appeals, our dedicated guide to getting hired as an FDE goes much deeper on the interview loop and the day-to-day.

Pay reality in India and the UK

Most published AI compensation data is US-centric, and the honest picture for India and the UK requires anchoring to sources that actually cover those markets. On levels.fyi, the self-reported median for a machine learning engineer in Bengaluru sits around ₹44.7 lakh, while the median for the newer "AI engineer" title is closer to ₹24.5 lakh — a gap that says as much about title inconsistency as about pay. In the UK, Indeed's mid-2026 data puts the average machine learning engineer at about £76,000 nationally, Glassdoor's London range runs roughly £52k–£107k for typical AI-engineer roles, and Lorien's 2026 UK survey describes a two-tier market with senior specialists at £100k–£150k+. Indian salary guides consistently report a 20–40% premium for demonstrated LLM, RAG or serving skills over generalist ML work.

The table below combines those anchors into indicative per-track bands. Nobody publishes clean per-specialisation salary surveys for India or the UK yet, so read these as directional — indicative ranges derived from levels.fyi, Indeed, Glassdoor and Lorien data plus the documented specialisation premium, July 2026 — not as quotable benchmarks.

Indicative ranges only, July 2026, from self-reported and survey data (levels.fyi, Indeed, Glassdoor, Lorien, plus US career-guide aggregations). Titles are inconsistent across companies; always benchmark the specific role.

Which track fits your background

The fastest pivots run along existing strengths. Every track rewards a different accumulated instinct, and the honest starting question is not "which pays most?" but "which one would I be 70% qualified for on day one?"

Then there is taste, which matters more than most people admit. If ambiguity energises you — half-specified problems, shifting requirements — agentic work will suit you; if it drains you, it will not. If you actively enjoy customer contact and the theatre of making something work in someone else's building, FDE is the only track where that is the whole job. If your satisfaction comes from systems that hum — latency graphs, utilisation curves — infrastructure will feel like home. And if your instinct on seeing a confident demo is "I don't believe you, show me the failure cases", you already think like an evals engineer. A track you are temperamentally suited to compounds; one you chose off a salary table does not.

Proof of work: what to build for each track

Every track has a canonical artefact — the single piece of evidence that a hiring rubric can score in minutes. Build one, properly, rather than five shallowly.

• Agents: one shipped agent with real users (even ten) and real observability — traces, failure rates, cost per task — plus a writeup that is honest about what broke. A polished demo with no failure data reads as marketing.
• Evals: an eval suite for a task you know well, plus an error-analysis writeup showing how it caught a regression a vibe-check would have missed. Our guide to building an evals portfolio covers the exact structure.
• Infrastructure: a serving benchmark writeup — one model, two serving stacks, measured throughput, latency and cost per million tokens, with configs published so someone can reproduce it.
• Context: a before/after case study — an agent or assistant that misbehaved, the context refactor you applied (curated files, retrieval changes, memory rules), and the measured delta in task success.
• FDE: a deployment case study with a business outcome — what you deployed, inside what constraints, and the number that moved for the customer. Even an internal-team "customer" counts if the constraint and the outcome are real.

The 90-day pivot plan

None of this requires quitting. The plan below assumes a full-time job and roughly six to eight focused hours a week — evenings and one weekend morning.

Weeks 1–2 — choose and audit. Use the background table above to shortlist two tracks, then read ten live job postings for each in your market (Bengaluru, Pune, London, Edinburgh — wherever you would actually apply). Write down the three requirements you already meet and the two you do not. Pick the track where that gap is smallest, not the one with the loudest salary headlines.

Weeks 3–6 — build the artefact. Build the canonical proof for your track from the previous section, scoped brutally: one agent, one eval suite, one benchmark — not a platform. If you can attach it to a real problem at your current employer, better still; an internal deployment is production evidence, and most managers will happily accept free automation.

Weeks 7–10 — write it up and publish. The writeup is not an afterthought; for evals and infra especially, the analysis is the artefact. Publish it, share it in the communities where your track's practitioners argue, and revise once based on the sharpest criticism you get.

Weeks 11–13 — re-position and apply. Retitle yourself around the track, put the artefact front and centre on your profile, and start applying against named-track postings — including internally. Moving from "generalist on the platform team" to "the person who owns our evals" inside your current company is the lowest-risk version of this pivot, and it sets up the seniority conversation covered in our AI engineer career ladder guide.

Ninety days does not make you staff-level in a new discipline. It makes you legible — a candidate with a named direction and one strong piece of evidence, which in 2026's market is more than most applicants bring.

Show the work where hiring managers look

Here is the quiet mechanical truth about how these roles get filled. Once a title becomes a track, hiring becomes rubric-driven: a screener looks for a named specialisation, shipped work that matches it, and a work history that makes the trajectory plausible. That is a specific artefact — and it is not a CV, which buries the signal, and not a social feed, which scrolls away. It is a profile: track stated plainly, projects listed with links, work history underneath.

That is exactly the shape of a Verified Builder profile on AI Tech Connect. You state what you build, list up to ten projects — including the artefact from your 90-day plan — and add your work history. People hiring across India and the UK browse the Builder directory to shortlist engineers by exactly the tracks in this article, and the profile costs nothing and takes about two minutes. No CV upload, no password, no gatekeeping — verification is the only filter.

One honest note on timing: the directory is young, and the earliest profiles carry a Founding Builder badge that later cohorts will not get. There is no countdown clock and no artificial quota drama — but being early in a directory that hiring managers are starting to browse is a genuine, compounding advantage, in the same way an early answer on a well-ranked forum thread keeps paying for years. If you have picked a track, the cheapest possible next step is making that choice visible.

Where this leaves you

The split of "AI engineer" into named tracks is good news dressed as pressure. It means you no longer have to be impossibly broad; you have to be legibly specific. Pick the track your background already leans toward, build the one canonical artefact, write it up in public, and put it somewhere the people hiring actually look. Do that over one disciplined quarter and you move from the crowded middle of a generic band toward the visible top of a specific one — in Bengaluru, in London, or anywhere between.