Agency Is the Dev Skill That Still Counts
Agents can write code all night. What a company can't get from an API is a person who owns the result when it breaks, and I think that's most of the job now.
An interview question
Picture a backend candidate who can recite Postgres internals from memory. Then someone asks: “Prod is throwing connection pool exhaustion errors and nobody knows why. What do you do in the first ten minutes?”
If the answer is “I’d wait for someone to tell me which service to check,” the Postgres trivia stops mattering. That one answer tells you more about how they’d do the job.
Volume was never the bottleneck
A company can run a fleet of coding agents that open pull requests overnight and never complain about boilerplate. Shipped code stopped being scarce a while ago.
It was never the constraint on a backend team anyway. The hard questions were about correctness under load. Will this migration run safely on a live table with 40 million rows? Will this retry logic turn into a self-inflicted DDoS? An agent will write the code with total confidence, and it won’t get paged or fired when the database falls over.
So the thing worth paying for is someone who answers for the endpoint when it fails. Writing it is the cheap part now. DigitalApplied’s 2026 hiring guide makes a similar point from the interviewer’s side: test whether a candidate can verify AI output instead of trusting it (DigitalApplied, 2026).
Agency and ownership
By agency I mean deciding what to do without being told. Ownership is still caring how the thing behaves after the ticket is closed. Those two together are what a company is buying when it hires you.
A low-agency engineer waits for a clearer ticket, hands it to an AI, and moves on once the PR merges. If that code melts the payments queue two weeks later, it’s somebody else’s problem. A high-agency engineer picks up the vague work and sticks around long enough to see whether it held up.
Do that for a few years and it turns into expertise, which a model can’t copy because it has never been woken up at 3am. Addy Osmani describes high-agency people as the ones who say “I’ll figure it out,” then keep going through obstacles and adapt as they go (Osmani, 2025).
The last 20 percent
My rough take: current AI tools get you about 80% of the way and stall on the rest. Concurrent writes, cascading failures, edge cases that only show up under real traffic. Those are judgment calls, and the model loses nothing by getting one wrong. If you can make those calls, AI makes you faster. If you can’t, it mostly helps you merge broken code sooner.
Hand that last stretch to the AI every time and your own judgment gets weaker. In an Anthropic study written up by byteiota, developers who learned a new library with AI help scored 17 points lower on a comprehension quiz than those who coded by hand (byteiota, 2026). Less practice means you’re worse at noticing when the model is confidently wrong.
What I’d actually do
- Review AI code like a stranger’s PR. If you can’t explain a line, don’t merge it.
- Take the vague tickets. Make a reasonable assumption, write it down, start building.
- Check on the code after it merges. Passing tests is where ownership starts, so go see how it behaves under real load.
- Have your own hypothesis before you ask the model. Use it as a second opinion.
- Write something without help now and then, so you can still catch the agent’s mistakes.
Knowledge used to hide it
For a long time deep technical knowledge and agency looked like the same trait. Getting good at complex systems meant grinding through ambiguity on your own, so the people who knew the most were usually also the ones who took initiative.
AI split them apart. Anyone can now produce fluent, plausible code without ever being the one who answers for it in production. Code got cheap and accountability didn’t.
Prompting better won’t fix this. Go back to the interview question: something’s broken, nobody has told you what to do. If your first move is to wait, I don’t know of a tool that helps with that.
Sources referenced:
Comments