Agentic Coding Cleared 'Can It Be Done' — The Hard Part Is What's Left

Agent harnesses paired with capable models have crossed the threshold of getting working code out the door, but the author argues that milestone is a small slice of the job. Drawing an analogy to welding things too heavy to move, he points out that whether software works is far less important than how it fits together — its seams, interfaces, debuggability, and long-term maintainability. Those qualities depend on subjective, experience-driven judgment about tradeoffs and abstractions, and that is exactly where today’s frontier models fall short.

The core limitation is that LLMs predict rather than reason; they are compressed human knowledge that can echo back reasoning only where it was already encoded (a point he ties to the ‘Illusion of Thinking’ research, and contrasts with emerging directions like JEPA and world models from Yann LeCun). The practical wins he sees come not from smarter prediction but from good tooling around the model: feeding it concise, well-timed context and wiring in deterministic validation that produces natural-language feedback the model can self-correct against. He notes economics are shifting too — cloud model business models look unsustainable while shrinking open-weight models increasingly run on beefy personal machines at close to comparable capability.

He also flags the security cost of tireless instruction-following: models can’t reliably tell good input from malicious, making prompt injection and Simon Willison’s ‘lethal trifecta’ foundational rather than incidental risks that alignment, sandboxes, and safety harnesses only partially contain. His conclusion is that as more people reach for agents to knock out ‘easy’ features, the durable, human skills — managing cognitive load, choosing the right abstractions, and deliberately reasoning about the seams — matter more, not less.