The Skill That Gets More Valuable When Agents Write the Draft

AI agents are getting good at producing the first version of almost anything - code, specs, campaign copy, analysis. The scarce skill is knowing, quickly and confidently, whether the output in front of you is right, and no longer just producing output.

The Idea: Specification Beats Production

Addy Osmani, who spent 14 years at Google working on Chrome DevTools and developer experience, discussed this shift in "From Chrome DevTools to AI Engineering, with Addy Osmani" on The Pragmatic Engineer. The pattern generalizes well beyond code.

Most professional work has two halves: deciding what should exist, and making it exist. Agents are reducing the cost of the second half. What they don't do is tell you the constraint you forgot, the edge case that matters to your specific customers, or that the request you made was the wrong request.

So the capabilities shifting toward value are the ones that sit on either side of generation: specification (being precise about intent, constraints, and success criteria before anything is produced) and evaluation (judging quality against those criteria fast). The capabilities losing relative value are the ones purely about throughput - writing boilerplate, assembling a standard doc, producing a first draft from a blank page.

This isn't "engineers become managers" or "PMs stop needing depth." It's the opposite. Judging an agent's output requires more domain depth than producing it did, because you're reviewing without the context you'd have built by doing the work yourself.

Real Example: Rewriting Your Growth Plan

Take a data scientist whose team now uses an agent for exploratory analysis. Here's the concrete shift, step by step:

1. Before: 60% of the week writing pandas transforms, 20% interpreting, 20% communicating results.
2. After: 15% prompting and reviewing transforms, 45% interpreting, 40% communicating and defending conclusions.
3. The trap: the agent produced a clean-looking cohort analysis that silently dropped users with null signup dates - 8% of the sample, skewed toward one acquisition channel.

Catching that required knowing the data's history, not knowing pandas. The person who spent the last two years only writing transforms wouldn't spot it. The person who spent that time arguing with stakeholders about what a "cohort" actually means would.

Apply the same audit to your own week. Estimate the share of your time an agent could plausibly absorb in 18 months. Then ask what you'd have left - and whether that remainder is something you're deliberately getting better at, or something you've been coasting on.

Key Takeaways

• Value is shifting from producing output to specifying intent and evaluating results - both require deeper domain knowledge, not less.
• Reviewing AI output is harder than it looks, because you lack the context you'd have gained by doing the work manually.
• Audit your week: what percentage is throughput an agent could absorb, and what's the judgment layer you're actively strengthening?

What's one task in your job where you can spot a wrong answer in under 30 seconds - and how did you build that instinct?

Sources referenced: The Pragmatic Engineer - "From Chrome DevTools to AI Engineering, with Addy Osmani"