AI will give you a confident answer. The hard part is knowing when it's wrong.
My approach to agentic engineering. The models are good enough now that producing the artifact is the easy part. The hard part, and the part I have built a practice around, is catching the confident wrong answer before it reaches a drawing. The full argument for why that is the skill that matters now is in Execution is the trivial part. What follows is how I actually work.
The trap
AI is very good at doing the wrong thing correctly.
It produces output that looks right, passes its own tests, and is quietly wrong, or that does nothing useful once it meets the real workflow. A tool that is confidently wrong is worse than no tool, because someone has to catch it, and by default nobody does. Teams that wire up AI themselves tend to find this out three or four projects in, when a number that looked fine turns out to have come from nowhere.
How I work
I treat a single AI answer as a draft from one fallible source, never as the result. The discipline is borrowed from how you already protect a sealed drawing.
The AI works in a sealed workshop, not loose in your systems.
The agent gets a contained workstation of its own: the tools it needs to do real work and to check its own results, and nothing else. It is walled off from your network, your files, and your operations, so even if a model goes off the rails it cannot reach anything that matters. You get the productivity without ever pointing an autonomous tool at your live systems.
A passing test is not the finish line.
AI-built work routinely passes its own tests and still does nothing in the real workflow. I check that it works in the actual flow, against the real persistence and command paths, before it counts as done.
The number comes from the source, not the model.
Where a value has to be right, it comes from the source document, the published table, or a federal data feed, with a citation. Never from the model's guess.
What that discipline catches in practice, and why it matters, is the subject of the post. Six features that passed every test and did nothing once the data was saved and reloaded. A plausible number the model invented. A path that data could have escaped through, which the tests and a first review both missed. The pattern is always the same: the win is not a clean first try, it is that the wrong answer got caught before it cost anything.
The honest part
Anyone can show you a demo where it worked. I would rather show you where it breaks and what I do about it. I am not a domain expert in everything I touch, and I do not need to be. The premise of AI uplift is that the model brings expertise I do not have. My job is to bring what the model cannot: a clear picture of what right looks like, the path to get there, and a working knowledge of where the model I handed the work to will mislead me.
How we would start
A free scoping call to find the one workflow worth fixing, a defined-scope pilot on one real project, and a production build only if the pilot earns it. Data handling is scoped to your confidentiality needs.