A method first, tools second

I started my career as a lean manufacturing engineer. The job was to walk a production line with a stopwatch, time every step and take out the ones that added nothing: waiting, moving things twice, fixing what the last station got wrong. The first rule was simple. You don't automate a step you should have removed.

AI hasn't changed that rule. Most rollouts that don't pay back have put a chat tool or an agent on top of a process that was already full of waste. Now people write the prompt, check the answer, fix the answer and paste it somewhere else. Automate a bad process and you get faster waste, plus someone checking the agent.

So every job starts with a map, not a model. Often the waste list alone saves more time than the first agent does.

Once the waste is gone, I split every remaining step three ways. It's the test I use to decide what an agent should do, and it answers the question most teams actually have: what can AI do without taking over?

Knowledge is looking things up, summarising, drafting and comparing. AI is already good at it, so it goes to AI. Skills are your team's repeatable know-how: the way you write a proposal, check an invoice or triage a ticket. Written down, they become skills an agent runs the same way every time. Wisdom is judgement, relationships and anything with real risk. It stays with your people.

The zone of proximal development is an idea from educational psychology: the gap between what someone can do alone and what they can do with the right support. Good teachers work in that gap. So do good agents.

With the right scaffolding (written skills, review gates and tests), an agent lets each person on your team work one level above where they are today. Seniors spend their time on design, review and decisions instead of routine work. Juniors and mid-levels take on work they'd otherwise wait a year for, with guardrails that catch mistakes early.

The result is the output of a bigger team for the cost of a smaller one, with your seniors multiplied and your juniors growing faster.

This is the engineering version of the same idea. Coding agents can write a feature in minutes, then someone has to review it. Pull requests get bigger, reviews get shallower, and the tests were often written by the agent that wrote the code, so they check what it did rather than what you meant.

Strategic programming splits the job cleanly. People own the architecture, the intent of each change and the tests that prove it. Agents write the code inside those lines, using skills your team has written down. Review gates check each change against its stated intent, not just its style.

The tests come first, and they come from a person. That one rule does more for trust in AI-written code than any choice of tool. Agentic Engineering