Vibe coding will deliver a wonderful proliferation of personalized software
Opinion For most of the last year, the phrase 'vibe coding' seemed more punchline than possibility. That outlook altered significantly over the last month after step-changes in quality mean vibe coding tools now generate code that’s good enough to rewrite expectations about how IT will operate before the end of this decade.
Until early November, none of the passes I've had at vibe coding prior yielded any results worth pursuing. Then I put Anthropic's Claude Code to work (courtesy of $300 in credit from Anthropic) transforming a prototyped Python data analysis project into a full-featured, professional-class tool. It took a few hours (and a fair bit of care and tending), leaving me with the impression that vibe coding had turned a corner into usefulness.
When Google dropped its latest Gemini 3 Pro model at the end of November, they left another gift under the tree: Antigravity, an integrated IDE and agentic coding tool that draws upon Google's recent acquihires from AI coding startup Windsurf. I fired up Antigravity with a thought about a bit of code I'd love to have – if I could have it with minimal effort. The answer came almost immediately: A VRML 1.0 browser.
VRML 1.0 came out in mid-1994, only to be quickly supplanted by the far superior and still reasonably-well supported standards VRML 2.0, VRML97, and X3D. That left a fair bit of early content stranded. VRML 1.0 translation tools have been lost in the mists of time. There's no way to load a VRML 1.0 world and view it.
So I fed Antigravity a copy of the VRML 1.0 spec, with the request that it create a VRML 1.0 browser for macOS, written in the Swift programming language, and using Apple's Metal 3D renderer. I didn't give it much more than that - but I did ask if it had any questions. It had a few, and then set to work.
Over the course of a Sunday afternoon, and Monday and Tuesday evenings, Antigravity churned out code, which I then plugged into Xcode, compiled, and ran.
While Antigravity worked, I functioned as its quality assurance assistant, because while Google’s tool can click its way around Chrome, if you're doing pure web work, it can't (yet) operate the macOS GUI. When compiles failed, I pasted those errors into Antigravity, which corrected its mistakes. When it built, I launched it, took screenshots, and fed those back in, with comments about the accuracy of the app's rendering.
By the end of Sunday night it was all working – sort of. Antigravity seemed quite pleased with itself, having plowed through the project checklist that it created at the outset and cleverly uses to keep itself on track when working on a large project. Yet you could have driven a planet between that working version and a fully compliant VRML 1.0 browser. Antigravity had done just enough work to render very simple VRML 1.0 files but had left vast portions of the specification unimplemented.
When I pointed that out, it drafted a new plan, and we spent Monday evening methodically implementing the features it didn’t code the first time. On Tuesday evening, I asked it to generate a conformance test for the app - which the app promptly failed. I fed that output back into Antigravity as a bug list.
That's when I had my penny-drop moment. Antigravity (and Claude Code) are quite powerful, but naïve. They don't know what they don't know. Unless a human gets in the agent harness with them, keeping them focused and aware, they tend to just sputter out. It felt as though I drove one of those massive dump trucks they use in Australia's mines: incredible power, that can easily tumble into a mineshaft - unless the human at the wheel remains continuously vigilant.
Skill at “steering” coding assistants may soon be the quality most sought after in software engineers, systems administrators and the like. It's a balance between a light touch and a firm focus.
After roughly eight hours of “steering” Antigravity, I had a VRML 1.0 browser for macOS - you can build it yourself from source. It might not be 100 percent compliant with the spec - that'll take a few more hours of testing. Yet it's already “good enough” to display all the examples from my 1995 book on VRML, and other bits of content I found online. To go from a spec and a goal to an app in eight hours feels like success. It would have taken me ten times longer to code this myself - and that's if I knew how to code in either Swift or Metal.
I wrote a few VRML 1.0 browsers thirty years ago, so I came to this with deep domain knowledge.
If I'd asked Antigravity to create a fault-tolerant database that would preserve transaction history across power failures, I'd have made a muddle of things because I don't know enough in that domain to steer the machinery. That means domain experts in software engineering and operations aren't going anywhere - but they will be getting a lot more productive.
Beyond these obvious wins, it's now possible to forecast a time – before the end of the decade – when these tools have been sufficiently "softened" to allow pretty much anyone within an organization to rapidly develop an app for a specific use case. That's an interesting bit of road directly in front of us: a Cambrian explosion of weird user-specific apps, designed by users for themselves. These tools make the best better, giving everyone else the benefits of write-once, run-once disposable software.
Vibe coding looks to be the gift that keeps on giving. Happy holidays! ®
