Why I Still Distrust Generic Agentic Coding

Agents can build a lot of code on top of one bad decision. So do humans.

I’ve seen this pattern throughout my career.

Someone starts without enough context, builds something that works, then asks for a review. Engineers with more context find fundamental flaws, and A LOT of code needs to change. Or the team accepts it because it works, and those decisions compound.

The tricky part is that “bad decision” is often subjective. It depends on experience, knowledge of the system, and opinions about tradeoffs.

Even after 22+ years in the industry, before a big change, I investigate, work out a proposed solution, and discuss it with people I work with. We refine the direction before I implement, check it again as the code takes shape, and review the result.

For agents, planning or spec-driven development is supposed to handle this. Many important decisions only become clear during implementation. Arguably, code is the best spec; a quick written specification won’t cover the important details. So agents do what they do best: fill the gaps.

Generic agentic coding seems to reproduce the workflow I already distrust: implement first, discover missing context during review, then untangle everything built on top of it.