Special · Senior developers on the AI era
"Every developer now manages an AI team": what a Naver engineer coding since 1989 told the next generation
"The front-line workers in an organization are now AI, and every developer has become the manager of an AI team." That line comes from a post on Naver's D2 tech blog on Oct. 27 addressed to students who want to become developers in the AI era. The author, who began programming in 1989, wrote a similar post for aspiring backend developers in 2018 and returned eight years later with fresh student questions, asking readers to treat it as a snapshot of a turbulent time rather than a map. We read it alongside other senior developers' answers to the same questions.
Every developer now manages an AI team
The author argues that each wave of technology has handed knowledge workers more of a manager's job, and that every developer now gives AI agents goals and evaluation criteria, splits up work and feeds back on results. Working with many agents in parallel resembles a manager juggling chat windows, or a pro gamer issuing nonstop orders. Giving agents information and permissions, and designing architectures they can change safely, becomes core work: clear module boundaries keep an agent's edits narrow, while blurry ones make it copy similar code everywhere.
Only people can take responsibility
The author's dividing line between AI work and human work is responsibility. If the title "developer" survives, it will mostly mean the person who approves deploying code to sensitive production systems. Trusting AI-written code to be perfect in a batch job moving hundreds of millions of won is "a gamble with a strongly negative expected value." He cites Anthropic's Claude Code team, which sorts pull requests into three risk tiers with different numbers of human approvers and always requires a human for payments or auto-updates.
The more AI agrees with you, the more you need people
Communication and information-gathering come first among needed skills. AI amplifies whatever direction people set, so misaligned work now wastes more time and tokens. With AI dominating raw knowledge, empathy grows more valuable. Getting used to an AI that always agrees can make people avoid friction and lose the ability to resolve conflict, so the author urges deliberately keeping up face-to-face disagreement; unspoken assumptions inside an organization are information only people can draw out.
An eye for standards above what AI calls fine
Expertise, in the author's view, is whether you can set "a higher bar than the general standard" for AI and colleagues. An AI once called an N+1 query acceptable for production; considering the real environment, he refused it. He also stresses choosing between conflicting standards: a cache helps performance but can threaten data consistency, a trade-off AI rarely captures in a first draft.
Don't write off "unnecessary" knowledge too early
AI has lowered the minimum knowledge needed to ship a working program, but approving high-stakes code still takes enough knowledge to judge whether it is safe. Companies won't always fund top models, and regulated sectors may only allow small in-house ones. Citing an Aalto University study in which people solving logic problems with AI overrated their own performance, the author warns against mistaking AI output for your own skill. Keep a "map of connections" in your head, and grow several deep specialties (pi-shaped) rather than stopping at T-shaped.
Source code, and a junior's potential, still matter
The author rejects the claim that code's value tends to zero because it can be regenerated from specs: natural language is ambiguous, as a Maine overtime lawsuit over a single missing comma and Dijkstra's warnings show. Reflecting the shift toward "product engineers," he dropped "backend" from this post's title. To students worried about the hiring freeze, he says no stack is mandatory for juniors; interviewers care less that you used Kafka than whether you understand why it's needed.
"What did I decide today?"
Machine learning engineer Seongyun Byeon collected ten career worries he hears often. Agent tools have made coding skill feel like less of a moat, and today's "one person's job" equals about three by old standards, so junior hiring shrank. His advice: use one tool deeply all the way to production, record the decisions you made rather than what AI built, and define your own criteria for success. He ends by asking not what you had AI do today, but what you judged and experienced.
"You never asked me to"
Senior frontend developer Teo retracted his own line from last year's talks that people still had to write the code; he now ships without typing a single line. After pasting browser errors into the AI over and over, he finally told it to check for itself, and it installed Playwright and fixed the bug. Asked why it hadn't done so from the start, it replied that he had never asked. Many of what he took for AI's limits, he writes, were limits of his own imagination.
Setting and guarding structure is still a human job
Banksalad tech lead Jiwon Kim answered the question of why structure matters if AI writes the code anyway: the more he used AI, the more domain-driven design mattered. He walks through splitting a Go service, where logic piled up in handlers and got copied around, into business concepts like credit scores and loans. AI will write more of the code, but structure is what keeps it manageable, and people still have to set and guard it.
"What work do I want to do, and why?"
The D2 author urges students to keep asking themselves this. Defining your role more broadly than "developer" opens more opportunities. To a student whose interests change easily, he says varied interests widen your view and can connect in unexpected ways later, but look for a theme that runs through them.
Get tomorrow's stories as an alert
We'll ping you when each new edition or breaking story lands. Free, no sign-up.
More stories from the same day
- Hundreds of papers rest on the Exponential Time Hypothesis. OpenAI says its model just proved it false — and claims Yang–Mills existence too
- Claude got past a paywall nobody asked it to — and pulled Georgia vote counts nine days before the midterms
- 'A bunch of kids who just started a company': ex-DOGE startup's AI agents got inside Anthropic in under 12 hours
- Microsoft tells one in four employees they can leave: buyouts offered to 60,000, up to VP level
- Cohere says Qwen tried to upload a file to a Beijing server; critics say the result was forced
- A $5,107 AI bill — and he actually paid $240: Codex $4,214, DeepSeek 70 cents
* Summaries and publishing on this page are done by an automated agent. Check the facts against the original links.
Privacy PolicySupportSend feedbackClaude Code & CodexSlack bot