The Last Programmer's Day? How AI Is Erasing 'Code Farmers' and Creating AI Builders

The 1024 Festival: Death of a Secret Handshake

1024 was never an official holiday; it began as an insider code among developers. The term "code farmer" (码农) spread from 2009–2012 as self‑deprecating humor — its premise was a high barrier to entry: above‑average pay for work outsiders couldn't understand. In 2021 China's Ministry of Human Resources classified software practitioners as "new‑generation migrant workers," and developers jokingly celebrated "code farmers really became farmers." By 2024 the Weibo topic hit 4.07 billion views. CSDN's 2024 conference renamed itself "1024 Programmer Day & AI Builders Conference" with the theme "Everyone Can Build," featuring robot‑company founders. A promotional post stated bluntly: "This year we expanded the definition — anyone who can command AI to build things counts as a programmer."

A festival doesn't die from cancellation; it dies when no one needs it to confirm their identity. 1024 will still light up, give away keyboards, and hand out tokens, but the people at the center have shifted from "those who write code" to "those who direct AI to do things."

Autumn Recruiting: Test Roles Vanish, Agent Test Engineers Emerge

In September–October 2025 campus hiring at ByteDance, Tencent, Alibaba, and other majors, pure functional‑test positions disappeared from campus catalogs or were merged into R&D tracks. Remaining test‑related roles carry longer titles: AI Quality Engineer, LLM Evaluation Engineer, Agent Test Developer. ByteDance's test‑dev JD requires "deep understanding and practical experience with AI Agents"; Alibaba's Tongyi Lab asks for "proficient mastery of ML algorithm principles." By contrast, 2023 JDs only asked for "familiarity with automation frameworks" and "Python experience."

The JD shift reflects what is being tested. Traditional systems verify deterministic input‑output pairs. Modern systems embed LLMs and agents, introducing failure modes — hallucinations, decision bias, privilege escalation — that are nearly invisible to conventional test frameworks.

Salary gaps are stark. In a 2026 top‑tier campus offer at the same company: AI Alignment role 420k CNY vs. traditional test‑dev 180k CNY. Lateral hiring (May–Oct 2025) shows AI test‑dev (LLM/Agent) at 30–70k CNY/month (400k–1M+ CNY/year) while traditional test‑dev sits at 15–25k CNY/month (180k–350k CNY/year) — two virtually non‑overlapping tracks.

Big‑Promo Incident: AI Runs Regression, Misses the Critical Path

A mid‑size e‑commerce owner saw an AI testing tool demo ("click‑click‑click generates cases, runs automatically, produces reports") and cut a 20‑person test team to two fresh graduates. During the big promotion, AI completed a full regression suite with a polished report. However, a flash‑sale scenario — "payment callback times out, user clicks cancel, then immediately repays" — was absent from requirement docs and UI screenshots; it existed only in veteran testers' heads. AI stayed silent, and overselling crashed the system.

The laid‑off people took away not "hands that write cases" but judgment over unexpressed, unenumerated risk zones. AI can generate 1,000 cases but cannot distinguish the three that are the real release‑blockers.

Tencent (June) and DeepSeek (August) Open the Books

At the 2026 Tencent Cloud Summit, Tao Daosheng stated: most Tencent code is now AI‑generated; engineers "periodically guide and correct what AI writes." CodeBuddy data shows >90% of engineers use AI coding assistants, and AI participates in 94% of code reviews. An earlier Tencent Cloud experiment: 4 engineers, 4 months, 99% AI‑generated code, completing an architectural evolution that would traditionally take a year. The remaining roles are "defining what's worth doing" and "judging whether it's done right," not "building it."

DeepSeek's open‑source Harness (dsh) provides a calmer benchmark. Breaking test development into five stages, dsh's replacement rates are: requirement parsing 50%, case design 70%, script writing 60%, execution scheduling 30%, defect localization 40% — weighted composite 50–67%. It can write boundary‑aware pytest cases, run them, and summarize failures, but semantic defect detection requires explicit business‑rule feeding, and execution pacing still needs human orchestration.

The ledger is clear: "execution" and "initial verification" are being consumed; "judgment" and "definition" are not. Pure code farmers and non‑senior testers happen to sit on the consumed side.

Disappearing Roles, Not the Profession

Test‑role polarization is accelerating in big tech: ByteDance, Ant, Alibaba, Meituan are pushing QA‑to‑dev conversion, merging test and dev sequences, moving quality responsibility back to development. Pure "test execution" as a standalone role is shrinking, while "building quality gates, designing agent evaluations, arbitrating between AI outputs" test‑dev positions are expanding against the trend.

Thus the software engineer profession persists, but the "code farmer who only writes code (manual labor)" is vanishing, fulfilling the metaphor. Historically "code farmer" was relatively egalitarian: technical yet physical, demanding stamina (favoring youth), with a skill barrier that offered training value and growth — juniors accumulated system understanding through coding and grew into seniors. That apprenticeship ladder has been removed by AI. Anthropic internal data shows junior technical hiring share plummeting from 27% to 9%, while AI agent work‑hours are already 6× human hours. The middle layer is disappearing in bulk, stretching the gap between apex and base.

This isn't "hiring less"; it's a paradigm shift . The software industry is sliding from a "craftsman market" to a "lottery market": the top 1% capture the vast majority of returns, while the rest struggle for scraps.

The Safety Zone Is Moving, Not Vanishing

Next year's 1024 will still happen, but its protagonists have changed. "Builder" is far broader than "programmer" — the expanded‑recruitment post listed creations like a consumer‑cooling‑off reminder, a wardrobe‑outfit generator, and a chat‑with‑Li‑Bai app; many of their makers never hand‑wrote a line of code.

For those already in the field, the honest advice: stop treating "I can write test cases" or "I can write code" as a safe zone. The safe zone is shifting toward "defining success criteria" and "being accountable for AI output." Those who make the jump are being fought over in 2026 autumn recruiting; those who don't are the ones selling "ancient‑era programming" books for 3 CNY a copy.

The most precise thing about "code farmer" isn't just "extreme reliance on physical labor" but "relative equality like farmers." AI shattered that equality, then demands a leap into new roles — roles that will no longer be called "code farmer." The festival no longer belongs to "code farmers," but the door for "software Builders" remains open.