Read enough Forward Deployed Engineer job descriptions and they blur into an impossible list: Python, distributed systems, LLMs, RAG, agents, evaluation, cloud, stakeholder management, travel, "comfort with ambiguity."
Nobody is exceptional at all of that. The useful way to read the role is as three layers, each doing a different job — and each ageing at a very different rate.
The three layers — and the human one is the moat
Layer 1 — Foundations
Python and software engineering. APIs, data structures, Git. Cloud and basic DevOps. Enough ML to reason about models.
This is the durable engineering base, and it is the layer people most often try to skip on the way to the interesting part. It is also where deployments actually fail. Not because someone misunderstood attention heads — because the code lived on one laptop, the API key was committed to the repo, or the container died on a dependency mismatch.
Two things deserve singling out.
Async is not optional. LLM calls take seconds. Twenty sequential calls take a minute; twenty concurrent calls take seconds. In a customer demo that difference is the whole impression.
"Enough ML" is a deliberately low bar. You need to reason about why a model hallucinates, what a context window costs you, and what temperature changes. You do not need to derive backpropagation. FDE hiring at the frontier labs is dominated by strong software engineers who moved into applied AI, not by PhDs.
Honest note on durability: this layer is commoditising fastest. AI coding assistants have compressed a great deal of routine implementation work. That does not make the layer optional — you cannot review what you cannot read — but raw coding throughput is a weakening differentiator.
Layer 2 — AI-specific
Prompt and context engineering. RAG and vector search. Agents, tools and MCP. Evaluation and LLMOps.
This is where current demand concentrates, and it is what most people mean by "AI engineering."
Prompt and context engineering has quietly shifted. The craft is less about clever phrasing and more about what goes into the context window and in what order — retrieved documents, tool outputs, conversation history, system rules, all within a budget.
RAG and vector search is the most common production pattern by a wide margin. The thing to understand is that roughly 80% of RAG failures trace to ingestion and chunking rather than the model. Teams that grasp this fix retrieval; teams that do not spend months tuning prompts against a broken retriever.
Agents, tools and MCP is where the role is heading. Worth knowing that a great many production "agents" are one model call, a tool and a retry — and that recognising when you do not need a framework is itself a competency.
Evaluation and LLMOps is the layer that separates people who can demo from people who can deploy. Without evaluation you cannot prove a change helped; without observability you cannot explain a regression. In a client engagement, "trust me, it's better" does not survive the first executive review.
Durability note: this layer is moving fast but the concepts are more stable than the tools. Frameworks churn annually; understanding retrieval quality, context budgeting and evaluation design transfers across all of them.
Layer 3 — Human skills
Problem framing and product sense. Communication and collaboration. Systems thinking and judgment. Fast, continuous learning.
This is the layer that looks softest on a CV and is the actual moat.
Problem framing. A client says "we want an AI chatbot." The valuable FDE works out that what they actually need is a better search over their policy documents, and that the chatbot would have failed. Deployments fail far more often from solving the wrong problem well than from solving the right problem badly.
Communication. The role means explaining a model limitation to a CTO in the morning and a business outcome to a CFO in the afternoon, without either conversation being a translation of the other. It also means telling your own product team a feature is not ready — and having the credibility to be believed.
Systems thinking and judgment. Knowing that an agent with write access to production is a different risk object from a read-only assistant. Knowing when 85% accuracy is fine and when it is unacceptable. This is largely unteachable from documentation and comes from having deployed things that broke.
Fast, continuous learning. The AI-specific layer will be substantially different in eighteen months. The skill is not knowing today's stack — it is the demonstrated ability to absorb the next one quickly.
Where to spend your time
If you are early: get the foundations genuinely solid. Everything else compounds off them and no amount of AI-specific knowledge compensates for not being able to ship.
If you are a working engineer moving into AI: the AI-specific layer is your gap, and it is a matter of months rather than years. See the sequential learning path.
If you already have both layers and are not getting FDE offers: it is almost certainly layer three. The strongest candidates are engineers with genuine customer-facing exposure — solution engineers who moved into applied AI, consultants who moved into implementation. If you have only ever worked behind a ticket queue, that is the gap to close, and you close it by getting in front of real users.
The uncomfortable summary
The three layers are not equally weighted, and the market weights them differently from how people study them.
Most self-directed learners spend nearly all their effort on layer two, because it is the most legible — there are courses, certificates and tutorials. Layer one is assumed. Layer three is treated as personality rather than skill.
But layer two is precisely the layer with the shortest half-life, and layer three is the one that determines whether a deployment succeeds. The engineers commanding the highest compensation are not the ones who know the most frameworks. They are the ones a customer's CTO trusts in a difficult meeting.
Our AI Engineering programme covers the foundations and applied-AI layers hands-on, using real deployment scenarios rather than toy examples — HRD Corp SBL-KHAS claimable for eligible Malaysian employers.
