The module addresses what remains irreducibly human in agentic engineering, and what erodes as models improve. Developers use AI in roughly 60% of their work, and yet they report being able to fully delegate only a small fraction of tasks. The contradiction resolves once effective AI collaboration is understood to require active human participation. Passive oversight will not produce the same result.

Developer work splits into three categories. Work that erodes: pure implementation of well-specified tasks, boilerplate generation, pattern-matching across large codebases, documentation of known behaviour. Work that is stable: debugging novel failures, evaluating ambiguous trade-offs, organisational navigation. Work that compounds: problem framing, system design, taste, the ability to evaluate agent output with genuine expertise, and the capacity to define what is worth building in the first place. The compounding category becomes more valuable as the eroding category shrinks.

Why some skills compound and others erode is, at root, a question Michael Polanyi described decades ago as tacit knowing. A practitioner knows more than they can tell. The parts of expertise that can be made explicit are the parts a model can learn from training corpora and reproduce. The tacit residue — the judgment and taste and sense of when to stop — is what stays scarce.

Four structural risks shadow the practitioner: erosion of programming expertise, black-box codebases, responsibility gaps, and ecosystem bias. In agentic engineering the human carries accountability for the output regardless of who or what produced it. The responsibility gap closes operationally rather than philosophically, through the audit trail, the specification, and the verification gates.

The compounding skills named here are precisely the skills the industry currently lacks. A 2026 systematic review reports four leading barriers to industrial adoption: trust and reliability (78% of studies), integration with legacy systems (64%), the absence of standardised evaluation methods (59%), and skills gaps in development teams (52%). The market is asking for the discipline this course teaches.