Learning Paths for Engineers in the Agentic Era

Around 90% of developers now use AI coding tools and roughly 41% of code is AI-generated. That does not make engineers obsolete, but it does move where their value sits — and a learning plan written for 2023 is now optimising for the part of the job that got cheapest.

What lost relative value

Not worthless — less differentiating. These are still worth knowing, and are no longer where a career separates:

• Syntax fluency in a specific language, and memorised standard-library surface.
• Boilerplate implementation of well-known patterns.
• Writing a first draft of anything with an established shape — CRUD endpoints, test scaffolding, config.
• Recalling API signatures, which was never a good proxy for skill and is now no proxy at all.

What compounds now

• Specification. Stating precisely what should be built, including what must not happen. Generation quality is bounded by specification quality, and vague specs now surface as expensive rework rather than as a conversation.
• Verification. Reading code you did not write and determining whether it is correct — under time pressure, at volume. This is the single most in-demand skill and the least taught.
• System design. Boundaries, contracts, failure modes. Machines generate within a design; they do not choose one.
• Debugging in production. Reasoning from symptom to cause in a live system remains stubbornly human.
• Evaluation. Defining what correct means for a non-deterministic system, and building the suite that enforces it. A genuinely new discipline.

A path that holds up

Four stages, each building on the last. The sequence matters more than the timeline, which varies by starting point:

• 1. Fundamentals that do not move. Data modelling, concurrency, networking, how databases actually behave. These have outlasted every framework cycle and will outlast this one.
• 2. Verification under load. Structured code reading, review practice on unfamiliar systems, tracing a bug from symptom to cause without the author present.
• 3. Specification and design. Writing specs precise enough to build from, ADRs, interface design, thinking in failure modes.
• 4. Applied AI engineering. Context engineering, retrieval design, evaluation, agent architecture, and the cost model underneath all of it.

Why stage 4 comes last

Applied AI skills are the fastest-decaying part of the path — specifics of a framework or a model's behaviour age in months. Built on stages 1-3, they are transferable, because someone who understands retrieval as an information problem will re-learn the tooling quickly. Built without them, they are trivia about a tool that will be replaced.

Making it real for a team

The failure mode of learning plans is that they are documents. What makes one stick is attaching it to work already happening: assign the ADR for a real decision, make the onboarding archaeology task the reading exercise, have the eval suite for a live feature be someone's growth project. Learning that produces a shipped artefact survives the quarter; learning that produces a completion certificate does not.