AI Agents Are Now Building Apps — and That Changes Everything

Chess pieces beside a board with artificial intelligence
Image: Remyvn via Wikimedia Commons (CC BY-SA 4.0)

There comes a moment every few years when the tools you use stop being assistants and start being authors. This week, that moment landed in a press release you might have scrolled past.

Snapchat announced Easy Lens CLI, a tool that lets Claude Code, Cursor, and Codex build, test, and publish AR Lenses for Snapchat — from a plain-language prompt, with no code written by the person clicking the buttons.

Read that again. The person describes the Lens. The machine writes the code, renders the preview, fixes its own mistakes, and pushes the finished effect to their phone. A game can go from an idea to a published Lens in a single day.

This is not a “no-code” release. It is something stranger: a no-human-in-the-review-loop release. And it changes what I think about when I say “the tools we build on.”

The numbers, and what they mean

Easy Lens CLI is free and runs on Mac, Windows, and Linux. Under the hood, the coding assistant pulls from more than 100 AR building blocks and 50 AI asset generators, then walks the full loop: build, test, publish. The assistant renders preview videos and sends them straight to the creator’s phone so they can check it before it goes live, fixes errors along the way, and ships updates in minutes.

Snap’s claim is deliberately modest: a game can go from an initial idea to a published Lens in a single day, without writing a line of code. Whether that is true in practice, I cannot say — I am not a Lens creator. But the direction is unambiguous. The barrier is no longer syntax; it is taste.

One number from the same announcement should give you pause: Snap’s Lens+ payout pool sat at $1.8 million last month, which works out to roughly $21.6 million a year if that pace holds. Add to that that developer tools like version control (Git-style) and paid “Digital Goods” are being opened to any Lens developer, and the picture is clear. Snap is no longer treating Lens building as a hobbyist plugin. It is treating it as an economy.

The trick is not new. The stage is.

Snap’s approach is not invented this week. Six months ago, the company applied the same agentic model to Specs, its AR glasses — and, yes, it also used Claude Code, Codex, and Cursor there. The new move is a step up in importance: instead of a head-worn prototype, the machine is now operating in the camera on every phone.

There is also the OpenAI layer. Snap has a ChatGPT Remote API so Lens developers can embed ChatGPT inside their Lenses, Lens Studio 5.0 Beta is shipping with a generative 3D face-mask generator, and the platform now counts 330,000 developers building 3.5 million Lenses that were viewed more than 3 trillion times last year. That is a bigger audience than most apps ever see.

What changed is that the creator stopped being the bottleneck between an idea and something a hundred million people can see.

Why “no-code” is a dangerous phrase

No-code has a shelf life. Every platform that promised it was eventually eaten by the next layer of tooling, and 2026 is the year the layer arrived.

The risk is not that the code will be wrong. It is that the assumptions will be wrong, and there will be nobody to notice. AR is unforgiving: a broken e-commerce page is annoying; a Lens that drops a 3D object, misreads a face, or leaves a game broken on the user’s screen is a small personal failure. The agent’s assumptions are invisible, and it can publish while you are still deciding whether the prompt was good enough.

That is the real change, and it is why I keep coming back to the idea in this piece: Your Agent Will Pick the Hackable One. The security boundary was never the model — it was the install. The same logic applies here: the model does not decide what is built; the loop, the tools, and the permissions around it do. If you do not care about security, if you do not vet the agent you are trusting, you are asking a learning machine to publish on your behalf into an audience you did not choose.

Vet it. Test it. Ship it to yourself first. The checklist is the old one — how to vet an AI coding agent before trusting it with your code.

Felix’s take: the human gate is thinning, but not gone

I have been a developer long enough to remember when “write a shell script” was treated as a legitimate skill and when “just use the cloud” was a punchline. The “prompt” is our new shell script, and the cloud is now a silent reviewer.

As someone who manages ICT systems and writes software, I have stopped believing that an unmoderated agent can ship into production on its own. Not yet. The tool is too fast, and the human is still the only one who knows what “good” looks like.

Chess and boxing come to mind here, because they are the two habits that shape how I think about this. In chess, the machine plays the board faster and makes fewer mistakes than I ever will. That is the advantage. But the machine cannot choose the opening, and if you walk into a struggle without knowing your plan, winning on speed does you no good. In boxing, resilience is not about hitting harder — it is about keeping your guard up and absorbing the first hit without going down. The AI tool is a faster hand. The guard is the review loop.

Here is the thing that should worry every Philippine developer and every ICT manager: the creator economy around Snap is young, and it is global. A schoolteacher in Manila, or a small-agency owner in Cebu, can now publish something that reaches millions, with no technical gate standing between them and the world. The upside is real — this is access, not exclusion. But the same door swings both ways. A bad prompt is an unfiltered business plan: it publishes an unfiltered product. The “easy” part is the risk.

Where this leaves the web’s spatial layer

For years, AR was a specialist’s hobby. It required lenses, face meshes, and a deep familiarity with Lens Studio. The Easy Lens apps already lowered that bar; now the LLMs inside them lower it again, because the assistant is what actually understands Lens Studio. The spatial web is no longer something you build — it is something an AI is willing to build for you.

That is a strange inversion of the “have AI write code” story everyone has been telling for two years — the AI agent stack is finally growing an operating system. The code is no longer the point. The outcome is. The user experience, the presence, the fit to the camera and the face — those are now the deliverables, and the model produces them at the speed of conversation.

If the model decides what gets built, then the next question is not technical. It is cultural. Who decides what matters? What happens to an audience when the creator never once looked at the rendering pipeline?

The bottom line

2026 is the year the writer moved from the keyboard to the prompt. The tool is faster, cheaper, and profoundly less accountable than a human editor.

The guard is still there. Whether we keep it in front of the machine, or spend the next year learning to trust it by default, is the decision we make now — not the one the press release hands us.

If you are a developer, ask your next agent to do something you would not personally ship. Then watch how it answers.

Filed under Tech & Gadgets
Last Update: November 19, 2026 by Felix AlterEgo
0 0 votes
Article Rating
Subscribe
Notify of
guest

This site uses Akismet to reduce spam. Learn how your comment data is processed.

0 Comments
Newest
Oldest Most Voted