Rigor, wrong aim
Holds the bar — but can't tell good architecture from bad.
Zaga · AI Lab — an engineering point of view
Keep the decisions. Hand production to the agent.
This is how we think about engineering now — the shift that changed the work, the eight principles that follow from it, and the engineer we hire, grow, and promote because of it.
When producing software became fast, the work moved up the stack. Less time goes into writing code. More goes into deciding what matters — and checking that the result is right.
Coding was never the easy part. What changed isn't the difficulty; it's the split.
The same tool cuts two ways. When building becomes cheap, more gets built — for better and for worse. The outcome depends less on the tool than on the discipline of the team using it.
AI unlocks speed. A playbook keeps the team on course.
"AI engineering" gets used for all three. We mean the first — nothing here requires the product to have AI in it.
One line holds all of it
Own the judgment, delegate the production.
Leverage moved upstream. Grounded practice is no longer about writing code; it's about directing, reviewing, and verifying what does.
Is this worth doing at all?
Read the need behind the request, and push back when scope or complexity outrun the value. Generation is cheap, so the default becomes to build more — knowing what to leave unbuilt is half the job.
How the whole fits together.
Hold the whole-system view and the concerns no single task raises. The model can draft architecture now: set the direction and judge what it proposes — never let it design unsupervised.
Break it into pieces you can steer and check.
Units small enough to steer and verify. Set the agent up to succeed before it begins — inputs, a clear definition of done, and the checks decided up front.
Give the model the right inputs.
Context is an asset you curate and reuse, not a throwaway prompt. Give each task exactly what it needs — relevant, complete — and keep the project's knowledge organized.
U1–U2 are the judgment band: what gets built. U3–U4 are the craft band: how the agent is set up to build it.
You answer for what ships.
"The model wrote it" is never an excuse — and "it works" is not the bar. You own whether the code is simple, reuses what's there, stays in bounds, and doesn't drift.
Confirm it's right with systems, not inspection.
Volume outruns line-by-line review, so verification becomes a system: risk-weighted spot-checks, tooling, monitoring — heavier where a miss is expensive. This is where the work bottlenecks now.
Stay sharp, stay hands-on.
Supervising well is not a step down from the craft — it takes more of it. Disengage and the decline is quiet: you won't notice until you can't debug what the agent can't.
Know what to use when — and where they fall short.
Know which tool to reach for, where the models are confidently wrong, and when not to reach for AI at all. Fluency is a moving target — spend the expensive models where they earn it.
G1 holds the standard. G2 proves it at volume. G3 and G4 keep the skills and the tools sharp.
A playbook provides structure, but people determine the outcome. These are the six traits we hire for, develop, and promote — the anatomy of an AI-fluent engineer.
Decides well — what's worth doing, what the real problem is, which way to go.
The eye for what good looks like, in systems and in prose.
Answers for the outcome — whoever, or whatever, typed it.
Does the unglamorous checking; holds the bar under pressure.
The standard. Cares that the thing is actually good, and won't let it slip.
Needs to know why and how; stays current as the tools move.
The tools will change. These traits matter regardless.
They look like one trait, but they aren't — you can have either without the other.
Holds the bar — but can't tell good architecture from bad.
Sees what good is — and won't let it slip.
Neither the eye nor the bar.
Sees good work — still ships slop under pressure.
Taste lets you see the gap between good and mediocre. Excellence is the refusal to ship the mediocre once you see it. Name both, and you know which is missing — a taste gap and a care gap take different coaching.
Taking the brief as given. Letting the agent design the architecture unsupervised. Trusting an all-passing test suite. Rubber-stamping what reads clean.
The root of the rest. Keep the decisions and hand over the work — not the other way around.
AI errors read clean, confident, and plausible — which is exactly why verifying, staying hands-on, and tool fluency are what catch what confidence hides.
Abdication — handing over the judgment, not just the production. The inverse of the one line.
A clean solution used to be cheap, fast, hard-to-fake evidence of the skills underneath — decomposition, debugging instinct, trade-off judgment. It isn't anymore.
A clean solution stood in for the judgment underneath. Cheap to run, and hard to fake.
When anyone can produce a clean solution, the solution no longer proves the skill.
Decomposition, trade-offs, product sense — plus the new skill: directing and distrusting agents.
This is what "AI-fluent" means at Zaga. Not that our engineers use the tools — everyone does. That they're vetted on the judgment the tools can't supply, and held to it on your team. We'd rather be honest about what's still an open question: the proxy broke, that much is clear. Whether the new formats fully predict on-the-job success is something the whole industry is still learning.
If you want engineers who bring judgment to multiply — not just hands to add — let's talk about what you're building.
Talk to us