If you read only the conclusion, the state of the frontend industry in 2026 fits in one line: front-end work was not killed off — it was taken apart.

The evidence does not come from panic. It comes from what the big companies wrote down this year. Read a handful of key posts and one deep-dive report side by side and you can see the major Chinese tech firms doing three things at the same time:

  1. Writing AI into the delivery process, and no longer treating it as a toy but as capacity;
  2. Turning team experience into assets that AI can reuse, moving it from “something a person remembers” to “rules an agent can read”;
  3. Erasing the “frontend / backend” job category from the org chart.

Those three are consecutive moves on the same line. Once AI pushes the cost of writing code close to zero, organisations recomputed why they should divide people by tech stack — and the answer came back: no longer needed.

This is not an AI news round-up. It answers one question only: who restructured the frontend role into what, and where does a person’s value now sit inside that structure?


1. Three things happening at the same time

Start with the hard numbers, because every judgement below rests on them.

Meituan’s Business R&D Platform put it plainly in a 7 May 2026 post, “Using an Agent-evaluation mindset to manage AI coding — the practice of an AI-driven refactor across 310,000 lines of code”: more than 90% of the team’s code is generated by AI; the system that carries this AI-coding practice grew from under 50,000 lines in June 2025 to 310,000 lines; the team ships 16 requirements per month on average (80% business requirements, 20% technical ones); and headcount tripled within a year.

Look at how those numbers relate to each other. Code volume grew more than sixfold, headcount grew threefold, and delivery stayed steady month over month. That is not people getting lazy. That is the output surface per human being stretched open.

LatePost laid out the cross-company picture in a 30 June 2026 report, “AI wrote 90% of the code: the painful moment for big-company programmers” (the link here is the NetEase syndication): Chinese big companies’ AI code rate was only 20–30% in the second half of last year, and reached 90% at the highest this year; Google announced in April that 75% of new code is written by AI; and according to Anthropic’s own report, as of May 2026, more than 80% of the code merged into its main repository was written and submitted independently by Claude — a number that was in the single digits in early 2025.

ByteDance’s technical VP Hong Dingkun gave a same-definition set of figures at a public event on 26 June 2026: over the past year ByteDance’s AI code contribution rate grew more than sixfold, AI-coding token consumption grew fivefold, and the AI code merge rate grew more than twofold. The merge rate is the key metric here — it measures the share of AI-written code that, after human correction, actually lands on the trunk. That number is doubling.

Then two signal-level events:

  • In the autumn of 2026, a screenshot of an internal notice began circulating across multiple accounts. Its content: “With the development of AI coding, the company has decided that from now on technical positions will no longer be divided by tech stack; all technical positions will uniformly be called Agent engineers.” The screenshot came from a Feishu message by former Alibaba P10 Bi Xuan, who has since left to found a startup. It is not an official Alibaba announcement, but during its circulation window it was discussed by a number of accounts.
  • On 12 November 2026, at Tencent’s 28th-anniversary company livestream, Ma Huateng raised AI-Native workforce restructuring (as relayed by the WeChat account AI Night Reading, 11.14).

Move the lens from data to people and it gets more concrete. The LatePost report names three:

  • A front-end engineer with ten years at Fliggy received a generous severance package on 11 June, and half the desks in the work area were empty;
  • A woman engineer at Shopee who had topped the performance ranking year after year was notified on 8 June;
  • Zhao Shu, a merchant-side front-end developer at a Shanghai company, was given a blunt reason for leaving: “the back end can now write front-end code by itself.” In her last month she had no work at all; afterwards she started learning Python to round out her full-stack profile, and when she interviewed, almost every question was about AI. She rated her chances at “70 per cent”.

There is also Gao Hong, a front-end engineer at Shopee who interviewed for an “agent engineer” role and came away with: “I kept finding that the interviewers often knew less than I did.”

So the question is no longer “will AI replace front-end engineers?” That question was settled in 2026 — partial replacement is happening, and faster than most people prepared for.

The real question is: if the job title itself is being erased, what exactly is being erased — the role, or the role as we imagined it?

My answer is the latter. The front-end role we imagine is a craft defined by slicing images / components / integration / styling. What the big companies are actually cancelling in 2026 is the division of labour by craft itself.


2. What gets cut first is the boilerplate, not the skill

First, one confusion worth clearing: the front-end work AI took over is not “front-end skill”. It is the slice of front-end work that is precisely specified, precisely accepted, and automatically detectable when wrong.

Alibaba’s Taobao Tmall Technology team summarised this in March 2026 in “AI coding in front-end practice: a retrospective”: AI is markedly efficient at boilerplate code, unit tests and UI components, but still limited on complex business scenarios. That matches what every team on the ground reports — what AI can write is “the things already spelled out in the requirements document”.

Break a front-end engineer’s day open and it looks roughly like this:

A day’s work Who does it Why
Slice the page from the design, tune spacing, write the static states AI takes over Precisely specified, enumerable styles, and a screenshot shows the mistake
Scaffold CRUD forms, list pages, modals AI takes over Highly repetitive patterns, mostly boilerplate
Write unit tests, add type annotations, generate API docs AI takes over There is an obvious acceptor — does the suite pass
Wire up data against the API doc, write mocks AI takes over Inputs and outputs are both written down
Debate with design about whether an interaction makes sense Human Requires judging intent, not translating intent
Pick between several options and state the trade-off Human Requires owning the consequence
After release, work out whether the bug is in the code or in the requirement Human Responsibility cannot be outsourced

That table lines up with Meituan’s numbers: more than 90% of code is AI-generated, yet 100% of requirements are still covered by developers doubling as testers (RD as QA), with a Human-in-the-loop testing SOP attached (Meituan A1, 7 May 2026).

In other words, fewer people produce code; the same number of people verify it.

One more thing is worth pulling out on its own. In the 2026 Enterprise AI Programming Practice Handbook published by ByteDance’s TRAE team in March 2026, the front-end section is not about “how to make the model write more like a human”. It is about Code Connect — establishing a mapping between design files and code — and about connecting external systems through the MCP protocol. The real engineering effort in front-end is therefore moving from “writing markup” to “feeding design intent, context and business rules to the agent”.

The model side is closing the same gap: Doubao 2.0 explicitly emphasises “enhanced front-end semantic understanding and CSS layout reasoning”. Once general-purpose models start optimising specifically for front-end semantics and layout, the moat around “being able to write it” is visibly shrinking.


3. Where the saved time goes: not leisure, but faster delivery expectations

This section deserves its own space, because it is where the lived experience of everyone working in this field actually comes from — and it is the exact opposite of “AI makes life easier”.

One line from the LatePost interviews recurred throughout:

“The efficiency gains from AI did not turn into leisure. Work that used to take ten days, the boss now expects delivered in one.”

And one more, colder:

“Being the one who uses AI to replace someone else is treated as the default reason you get to stay.”

Understanding the mechanism behind that second line matters more than remembering any single percentage.

The unit of assessment did not change; the amount of work per unit time did.

What you see is an entire apparatus of institutionalised behaviour:

  • One ByteDance team put the AI code rate into its OKR and displays it as a rolling counter on a website;
  • At Alibaba’s Taobao Tmall Group, senior management pushed a unified dashboard where every individual’s AI usage data is visible to everyone else;
  • Meituan asked every team to map out “which work processes can be standardised and which steps can be distilled into skills”;
  • According to a Meituan veteran quoted in the LatePost report, after management pushed this top-down, the company’s AI code rate went from 10% to 70–80%, and a lot of people got it running within two weeks. Wang Xing’s remark after meeting AI people is quoted repeatedly: “The water has risen… whether you are struggling in it or not, you all have to learn to swim.”

So the real feeling is not “there is less work”. It is that far more has been packed into the same working hours. A front-end developer used to ship two pages a week; now the expectation is two pages, plus a dashboard, plus a regression pass.

That is also why the efficiency dividend never became a benefit: the dividend was collected by the organisation, and the dividend was never the employee’s to distribute.

One more detail, from He Chuan, a Coding Agent developer inside Alibaba, who describes his situation as an “internal battle royale” in which P7s, P8s and P9s each fight their own tier. He says: “Everyone wants a place of their own in the AI era, but the likeliest outcome is that you don’t have one.”

It sounds pessimistic, but it describes a genuinely new competitive mechanism accurately: the axis of evaluation has moved from “how much did you write” to “how much did you help the team produce, and how much can the result be trusted”.


4. Three turning points of the role

The previous section was phenomena. This one is structure. Here is what the big companies’ own material supports, sorted into three turning points.

Turning point one: the tech-stack category is abolished — the role shifts from a division of labour to a unit of responsibility

The most direct evidence is the circulating screenshot: “From now on technical positions will no longer be divided by tech stack; all technical positions will uniformly be called Agent engineers.”

In the same period, a number of workplace-media accounts carried the same story: Ant, JD, Meituan and Dewu had, one by one, dissolved or merged their standalone front-end departments; JD removed the “front-end engineer” hiring keyword; Meituan folded its front-end team into the back end. Rigour requires precise labelling here:

  • According to the essay “Code Traitors and the Industrial Revolution”, Dewu’s dissolution of its independent front-end department happened in March 2025; November 2026 is the renewed reaction to it, and Dewu has officially responded;
  • The claim that “Ant removed the job title ‘front-end engineer’ from the organisation in July (2026-07)” comes from workplace media, not an official announcement;
  • These are rumours and retellings, and must not be written as established fact.

But the direction is unambiguous. The LatePost interviews put it more precisely:

“Role boundaries are dissolving — front end is being merged into back end, and newly posted roles are simply renamed ‘full stack’.”

So why would an organisation abolish division by tech stack?

The reason for dividing labour was never the technology itself. It was to reduce communication cost and learning cost. Letting a front-end developer write a back-end endpoint used to cost “three months of learning plus cross-stack communication friction”; AI has pushed that migration cost close to zero — which is precisely what the circulating screenshot means by “a front-end engineer no longer needs three months to learn how to write a back-end endpoint”.

Once migration cost approaches zero, the benefit of division (saved communication and learning cost) disappears while its friction cost (hand-offs, waiting, boundary disputes) remains or even grows. So organisations naturally change the axis of division: from “by language and directory” to “by business domain and delivery outcome”.

Going from front end to full stack is therefore not “learning one more language”. It is redefining the role from a list of skills to a list of responsibilities. You are no longer on this team because you know Vue; you are here because you own the outcome of this business domain.

Turning point two: code review and acceptance become the new bottleneck — human value moves from “writing” to “proving”

This is the most important of the three turning points, and the one most easily missed.

Meituan’s Basic Technology department wrote three sentences in its 29 October 2026 post, “Using AI coding in a system that cannot afford mistakes: practice and boundaries in distributed storage”, each of which can stand as a conclusion on its own:

“AI moved the bottleneck from ‘writing’ to ‘verification’.”

“AI’s mistakes are not in the code, but outside the code.”

“Fluency is not evidence. Consistency is not correctness. Green is not resolution.”

“Fluency is not evidence” is worth reading twice by anyone who writes code. A model that walked a flow end to end, a suite that went from red to green, code that looks tidy — none of these constitute correctness. It may in fact be wrong.

What “verification” now looks like in practice, according to that same Meituan post, is worth copying: a whole-repository review surfaced 117 suspected issues, of which 36 were confirmed by red tests, 3 were disproven, 41 confirmed by reading source, and 37 remained unverified — note that last column: the 37 unresolved cases were not counted as results. Another, harder set of figures: 678,000 lines of WARN logs were scanned over 11 days across 34 call sites, and the 5 defects that emerged were all real, with zero false positives. The CMake overhaul covered 31 build files.

That is what “verification” looks like now: not a glance at a diff, but the design of an evidence chain that can force an error into the open.

Which is why the conclusion in Meituan’s A1 post is the centre of the whole article:

“When 90% of code is generated by AI, the focus of the team’s work should shift from ‘writing code’ to ‘designing and maintaining an engineering environment in which AI can reliably produce code’.”

And the nature of code review changes with it: the value of human review shifts from “did you write it correctly” to “are we solving the right problem under the right constraints”. The first is error-checking; the second is boundary-setting.

Add Meituan’s A2 principle: AI debugging can only act as a “hypothesis generator”, and root-cause attribution must be done by a human. And the layered authorisation path they designed: read-only analysis → controlled modification → controlled execution → human decision — the further right you go, the heavier the responsibility, and the less it can be outsourced.

The conclusion is blunt: the most expensive skill in the AI era is not writing code. It is signing off on what the AI produced.

Turning point three: experience becomes an organisational asset

Once the job categories are abolished, the thing organisations fear most is obvious: experience leaves with the people who hold it.

Ten years of front-end experience means knowing every trap in the business domain: which field turns out to be dirty, which flow blew up three years ago, which component change detonates three other pages. All of that used to live inside a senior colleague’s head; the moment roles merge and seniors leave, the capability leaves with them.

So the third thing the big companies did this year was turn that experience into something an agent can read:

  • Meituan turned conventions into always-on AI Rules + Skills + pre-review checks, and framed the whole method as a move from “human-to-human alignment” to “human-to-machine alignment” (A1);
  • Alibaba: Tmall built an “AI full-stack R&D knowledge base”, distilling business documentation and coding standards into assets the whole team can reuse, to solve the problem that general-purpose models do not understand business logic (Taobao Tmall Technology, March 2026); a second post the same month, “AI coding: from tool efficiency to paradigm shift — what are we still missing?”, names the bottleneck precisely: “duplicated knowledge construction, and a hard time reducing cost”;
  • ByteDance: the TRAE handbook breaks down context engineering, with the front-end section using MCP to reach external systems and Code Connect to map design files to code;
  • Tencent: on 11 March 2026 it launched SkillHub, “an AI skills community optimised for users in China” — productising skill packages and SOPs.

Even the working rhythm of front-end teams themselves has changed. Shouqianba’s front-end team summarised a usage SOP adapted to the front-end rhythm in “AI coding practice: from blind use to control”. Meanwhile, Ctrip’s flight-tickets front-end team wrote “Why does your AI coding get stuck at Copilot?” — and the answer is: what blocks you is almost always context that has not been set up properly.

Neither of those two is an official big-company document, but together they say more than the official ones would: when “configuring the agent’s context” becomes a daily chore, the core work of a front-end developer has already shifted from writing code to writing configuration.

And with that, the value of the veteran front-end developer changes too: from “writes fast” to “writes the rules well”.

In the past your experience lived in your head. Now it has to be written as the agent’s Rule.


5. Four big companies, four answers

The same problem, four different answers. Put together, they are complete.

Company What it did this year What it means for front-end roles
Meituan 90% AI-generated code + an agent-evaluation system + layered authorisation (read-only analysis → controlled modification → controlled execution → human decision) + RD as QA (Meituan A1, A2) The role shifts from “delivering code” to “designing and maintaining an engineering environment in which AI can reliably produce code”; the AI code rate goes straight into performance review
Alibaba Tmall’s AI full-stack delivery practice + AI full-stack R&D knowledge base + Qoder as an AI-native IDE (expert-team mode, “Leading Technology Award” at the 2026 World Internet Conference; the company says it serves 8 million users); Qianwen Office public beta on 2026-08-03 The tool layer makes “one person, many stacks” the default; the asset layer turns business knowledge into shared property; management drives individual productivity through a unified dashboard
ByteDance TRAE’s 2026 Enterprise AI Programming Practice Handbook (context engineering; MCP and Code Connect for front end); Doubao 2.0 strengthening front-end semantic understanding and CSS layout reasoning; Hong Dingkun disclosing AI code contribution up more than sixfold, tokens up fivefold, merge rate up more than twofold; 37,000 developers in the TRAE AI Creativity Contest; a 2026-11-03 internal rule restricting external AI tools from internal codebases Models and tools are deliberately closing front-end’s weak spots, while AI usage is simultaneously brought into performance review and governance boundaries
Tencent Two product lines — WorkBuddy (an all-scenario AI office agent) and CodeBuddy (for software development), with Liu Yi as Tencent Cloud VP; 2026 Q1 active-user retention above 60% and paying-user retention above 80%; China Academy of Information and Communications Technology FDE validation in November 2026 plus an ecosystem-partner programme; the SkillHub skills community from 2026-03-11; CodeBuddy CN and Alibaba Cloud Bailian Coding Plan both moving to paid tiers From “free tools for trials” to “priced per headcount, plus third-party certification, plus ecosystem revenue share” — AI capability now drives headcount and budget directly

Each of the four carries a different emphasis.

Meituan has gone deepest on “how do you manage code that AI wrote”: from the agent-evaluation mindset to layered authorisation, from “aligning people” to “aligning machines”. Its definition of the role is remarkably complete — what you design is not code, but an engineering environment in which AI can reliably produce code.

Alibaba is pushing three lines at once: tools (Qoder as an AI-native IDE rather than “an IDE plus a plugin”), knowledge (Tmall’s AI full-stack R&D knowledge base), and management (Taobao Tmall Group’s unified dashboard where every individual’s AI usage data is visible to everyone). Qoder’s expert-team mode has a vivid pitch: “one sentence summons 13 ‘programmers’”.

ByteDance requires reading three distinct voices together. The TRAE handbook is “we want to help you build front end faster”. Hong Dingkun’s figures are “we have already been doing this ourselves”. And the internal rule of 3 November 2026 is “external tools are not allowed near our code”. All three together are a complete organisational answer: accelerate adoption while tightening the boundary.

Tencent concentrates on the product side: turning AI programming into a commercial product with retention rates, payment, third-party certification and an ecosystem. One detail is worth noting: on desktop monthly visits, Tencent’s WorkBuddy leads at 20.97 million, ByteDance’s TRAE IDE at 12.79 million, and Alibaba’s QoderWork at 7.88 million (The Paper, 2026-08-10). The AI-native IDE is competing for the desktop entry point — and where the entry point sits eventually decides headcount.

As for Tencent’s own front-end teams describing the change in their own words — no official 2026 long-form post has been found. The pieces that exist on the community site, “Front end will not be killed by AI” and “The impact of AI on large front-end projects — resisting with Big Front-end++”, are frontline arguments from October to December 2025, and should not be packaged as “what Tencent said this year”. That said, one judgement from them is worth keeping: “AI does not destroy a profession; it makes the profession more elite” — and “the front-end developers who will be valuable in the future are not the ones who can write code, but the ones who know how to create vibe with code.”


6. The compressed part: not gone, pushed down

Here is a counter-intuitive conclusion: the parts of the front-end job that got compressed away have not vanished into thin air. They have been pushed down into three places.

  • Into the model: boilerplate code, UI components, unit tests, API documentation;
  • Into the toolchain: scaffolding, CI, build systems (Meituan’s A2 CMake overhaul covered 31 build files);
  • Into the process: conventions, acceptance, audit, permissions (Meituan’s layered authorisation turns “am I allowed to change this” into a rule of the process).

Being pushed down is not the same as disappearing. Boilerplate code went from “typed by a human” to “generated by a rule”, but the rule is still written by a human, and a wrong rule is still a human’s problem. So the new skill is “writing rules”, not “writing code”.

One angle here deserves special attention: the first to be compressed are the people at the very bottom, with the least bargaining power. New Weekly’s “A portrait of coders under the AI shock: the outsourced labour pool shrinks sharply, and writing basic code is no longer a living” (syndicated by Sina Finance, 2026-11-05) documents exactly this — the pool of outsourced and basic-skill roles is contracting fast. It is both the most direct victim of this “pushing down” and a warning: the order of descent is fixed; people near the top of the pyramid are not affected last, but they are not immune either.

On the hiring market, one self-media account offers a frightening pair of numbers: front-end job openings down 51% over the past year; another claims a 40% salary premium for “front-end + AI” skills. Both come from self-media, and I could not find a cross-checkable primary source, so treat them as “somebody says so” rather than data: the trend is credible; the numbers are not yet verified.


7. Four objections, answered head on

An article about “the job has been rewritten” is vulnerable to four objections. Rather than dodge them, take them one at a time.

1. “Contracting jobs ≠ AI substitution”

This criticism is entirely fair, and the record contains explicit counter-examples: reports on ByteDance’s Wuhan team note that the reduction was mainly business-line contraction, not direct AI replacement.

More importantly, the LatePost report contains a qualifier that must be quoted: not a single Chinese company has publicly announced layoffs caused by AI substitution. The reduction has happened through flexible means — not hiring, not renewing contracts, adjusting the business. The timeline in that report is presented the same way: 630 (30 June) became a shared Last Day for a group of people — Fliggy on 11 June, Shopee on 8 June, and 30% of one Meituan department on 17 June.

So this article cannot use “laid off” to describe big-company front-end work either. This article is about structural change, not about headcount reduction. Whether the total is falling and how the structure is changing are two questions that have to be answered separately.

2. “All your data comes from self-media”

Correct. Which is why, in the body of this article, hard numbers only come from sources you can open and check: Meituan’s two official posts (310,000 lines; 90%+; 117 → 36/3/41/37; 678,000 lines of WARN logs; 31 build files), the LatePost deep dive, ByteDance’s Hong Dingkun speaking on the record in Securities Times, and The Paper’s monthly-visit comparison. Any figure from self-media is either labelled “one account puts it at X” or left out.

If a number has no source, it does not belong in this article.

3. “Going full stack just means front-end engineers get more work”

This worry is real. The Tencent Cloud community post “The impact of AI on large front-end projects — resisting with Big Front-end++” puts it bluntly: “Almost nobody is hiring front end now, because back end using AI-assisted development can basically meet the need.”

But the LatePost line about efficiency gains never turning into leisure makes the point sharper: if going full stack only means taking on more work, that is a failed version of going full stack. Success has two markers — business throughput per head rises, and judgement does not get outsourced. If the second condition fails, the first one is worth much less: you are not simply doing more, you have turned yourself into a cheaper execution unit.

Meituan’s A1 conclusion works as the standard: the centre of gravity moves to “designing and maintaining an engineering environment in which AI can reliably produce code”. That role is not “more work”; it is “being responsible for the work being less wrong”.

4. “The more code AI writes, the easier life gets”

Meituan’s A2 answers this in three sentences: “Fluency is not evidence, consistency is not correctness, green is not resolution”, and “AI’s mistakes are not in the code, but outside the code”.

More output does not mean less risk. Verification cost is the new scarce resource. A team that writes ten times the code but cannot say why any of it is correct is in far more danger than a team that writes less and can account for every line.

One more piece of positive evidence belongs here: the circulating screenshot’s line — “a front-end engineer no longer needs three months to learn how to write a back-end endpoint”. Falling migration cost is AI’s most direct contribution to the individual. It turns “become full stack” from something that took three years of gritted teeth into something you can learn while you work.


8. When the job title is gone, where do you stand

If you want an actionable framework, I would split “will front end disappear?” into three questions you can actually answer:

  1. Does your team still have a job title? (organisational layer: Agent engineer / full stack / split by business domain)
  2. Are you still delivering “code”? (artefact layer: specifications, rules, acceptance criteria)
  3. Who is accountable when it goes wrong? (responsibility layer: the person who signs cannot be outsourced)

The further right your answer sits on each question, the more secure the role. If a job title is already being renamed at layer 1, no longer delivers code at layer 2, and has no accountable person at layer 3, then it is close to being compressed.

Which yields four action items, each tied back to the material above:

One: productise your judgement. Learn to write always-on AI Rules, Skills and acceptance scripts, so that conventions move from “suggestions in a document” to “constraints the agent executes every time”. This is Meituan’s “human-to-machine alignment” (A1) — if your experience is not written down, it does not exist.

Two: treat verification as the main event. Build the ability to attach an evidence chain to AI output: red tests, log audits, regressions, reproducible defect descriptions. This is the 117 → 36 sequence in Meituan’s A2 — people who can falsify are scarcer than people who can write.

Three: move towards the business domain. Generic syntax is a public good; business rules are not. The value of Tmall’s knowledge base comes entirely from distilled business knowledge. If your experience cannot be written as a business rule, it is still a personal trick, and it will disappear with you.

Four: hold on to taste. This is the layer models are worst at. The conclusion from the Tencent Cloud community post is well put: “the front-end developers who will be valuable in the future are not the ones who can write code, but the ones who know how to create vibe with code.” Vibe is not mysticism; it is the ability to judge, among several options that all “work”, the one that feels better to people — and that takes taste, intuition and a great deal of practice.

And one last note, from the LatePost interviews, which I think matters most: turn “I can hold on alone” into “I can get the whole team through it”. The scarce thing in the AI era is not individual throughput; it is people who can define constraints, distil rules, and carry the final judgement.


Closing thoughts

Back to the three signals from the opening: Meituan’s 90%, the circulating “Agent engineer” screenshot from Alibaba, Tencent’s AI-Native session. They point at the same thing: technical roles are being re-sorted from “classified by skill” to “classified by responsibility”. Under that new classification, “front end” increasingly becomes a label like “fluent in a particular framework” — it describes where you came from, not what you are responsible for.

So the question worth asking is not “does front end still exist”. It is: under the new division of responsibility, is the piece you own something nobody else can take?

If the answer is “what I do is write the things AI can now write faster than I can”, then it does not matter what the role is called — the risk is identical either way. If, on the other hand, you hold the business rules, the evidence chain, and the final judgement, then the title is irrelevant: an organisation ultimately needs someone to be accountable for the outcome, and that someone has to be a person.

The role shrinks, the value moves. From front-end engineer to Agent engineer, what changes is not the name — it is the range of what you can put your signature on.


Source card

The sources cited here, listed by how much weight they can bear:

Official big-company engineering blogs (readable in full)

Deep reporting and public statements

Frontline community arguments (originally published in late 2025, not this year’s posts)

On citation discipline: the circulating internal screenshot and all self-media figures in this article are labelled as rumours, reports, or self-media estimates; any percentage that could not be traced back to a primary source has been left out.

评论