What it is
Agentic work changes what skill to train for. Historically, writing code by hand was how practitioners trained their judgment — the act of implementation built the mental models. When the machine writes most of the implementation, that training path narrows, and the relevant skill becomes learning to steer, decompose, inspect, validate, and govern agentic workflows rather than to type the solution. The new literacy is not prompting cleverness — not a bag of phrasings that coax a model — it is operating: the durable competence of running a production system made of agents. Concretely, training for it means teaching people how to express intent precisely, how to ask clarifying questions before work begins, how to build acceptance criteria that define done, how to inspect generated work rather than trust its confident appearance, how to run adversarial review, how to use agent-shaped tools, how to manage multiple sessions at once, how to avoid poisoning context, how to recognize when to stop the machine, and how to turn repeated practice into reusable skills or patterns. These are operating skills, transferable across tasks and models, and they are what separates someone who supervises a fleet of agents effectively from someone who merely issues prompts and hopes.
When to reach for it
- When designing training or onboarding for an agentic team: build the curriculum around operating skills — intent, decomposition, inspection, validation, governance — not around prompt phrasings.
- When a practitioner can get an agent to produce output but cannot reliably tell whether it is right, invest in inspection and acceptance-criteria skills before more prompting tricks.
- When someone is managing several agents and dropping work, teach session management and attention discipline as first-class skills.
- When repeated manual steering keeps recurring, train people to capture it as a reusable skill or pattern instead of re-improvising each time.
- When evaluating or hiring for agentic roles, assess operating competence — can they steer, inspect, and stop the machine — rather than prompt fluency alone.
When NOT to
- Do not dismiss prompting entirely; clear expression of intent is part of operating — the point is that prompting alone is an incomplete literacy, not that it is worthless.
- Do not push operating skills onto people who still need to build foundational domain or engineering judgment first; operating amplifies judgment but does not manufacture it.
- This is not an argument that hand-writing code is now useless for learning — deliberate hands-on practice still builds judgment that pure operating does not.
- Avoid turning "operating" into another set of memorized tricks; the literacy is transferable competence, not a new checklist to game.
Exemplars
- Andrej Karpathy, "Software 2.0" — https://karpathy.medium.com/software-2-0-a64152b37c35 — classic antecedent: the shift from writing explicit code to specifying desired behavior anticipates the deeper shift this pattern names — from implementing to operating.
- Simon Willison, "Vibe engineering" — https://simonwillison.net/2025/Oct/7/vibe-engineering/ — formalizes: distinguishes rigorous AI-assisted engineering (leaning on testing, planning, code review, and an updated sense of estimation) from mere prompting; argues that AI tools "amplify existing expertise" — operating judgment is what gets amplified, not what gets replaced.