An Agentic Bake-off: Antigravity vs Claude Code
I replaced the site I hand-built in 2018 with one built almost entirely by coding agents — running Google's Antigravity and Anthropic's Claude Code at the same job to see where they differ.
The site you're reading used to be something I hand-built in 2018 — a
student portfolio template with my name on it, lovingly assembled the
traditional way: one file at a time, at one o'clock in the morning. The
repository is still called node-calumbell, a small fossil from that era.
This version was built differently. Almost none of it was typed by me. Over a few evenings I rebuilt the whole thing — application, infrastructure, deployment pipeline, DNS migration, and most of the design — using coding agents, and because I couldn't resist turning it into an experiment, I ran two of them at the same job: Google's Antigravity, driven by Gemini, and Anthropic's Claude Code. Same brief, same repository, same opinionated-but-vague client (me).
What they were replacing
The 2018 site was the kind of thing every engineer has in a drawer: an Express app serving a JSON résumé, deployed by hand, quietly rotting. The brief for its replacement was real work, not a toy: a Next.js site with MDX publishing, a password-gated CV served from private storage, a contact form wired to a transactional email service, Terraform for all of it on Azure, GitHub Actions deploying containers on every push, and a DNS migration off the legacy setup without breaking the email domain my actual life runs through.
Both agents handled that structural layer well, and honestly, similarly. Routes, API handlers, Terraform resources, CI workflows — this is bread and butter now. If your mental model of coding agents is still "autocomplete with confidence," the accurate update is: the entire boring 80% of a project like this is done, reviewably and with working CI, faster than I could have scaffolded it myself. On that 80%, I'd struggle to tell you which agent built which half.
Where they separated
Two differences mattered.
Cost. Gemini operated at a noticeably lower cost point for comparable structural output. If the work is well-specified — build this route, write this pipeline, migrate these records — that economics is hard to argue with, and at enterprise scale it's the kind of difference that shows up in real budgets.
The last 20%. Claude Code was a visible step up when the work stopped being specification and started being design. The clearest example is the thing you may not have noticed yet: move your cursor around the homepage. There's an engineering schematic of my career hidden behind the page — faintly visible in one corner, revealed by the pointer like a torch over a blueprint. That feature started as one vague sentence from me ("something subtle but cinematic"). What came back wasn't just working code: it was a design argument — where the drawing's density should live so text stays readable, why touch devices should get nothing rather than a broken imitation, a masked SVG approach I'd have needed a day to land on myself. When labels overflowed their boxes, it measured every one programmatically rather than eyeballing a screenshot. When I supplied a new headshot, it stripped the GPS coordinates out of the EXIF before publishing without being asked. At one point it corrected the site's copy against my own CV — the copy said I'd held a job title I never actually held.
What was left for the human
More than the hype suggests; less than my ego would like. My contributions, honestly listed: taste ("this is too boisterous, make it more humble"), accuracy ("that number is wrong, pull the post"), judgement about what should exist at all, and the occasional correction when confident output drifted from the truth. That last one matters — an agent will build a wrong thing beautifully if you let it.
But that's the interesting shift. The engineering constraint on a project like this used to be typing capacity and patience. Now it's the quality of your judgement, applied at conversation speed. The bar for "small team" just dropped again — this site is a one-person team, and the one person mostly reviewed.
I've spent years telling people that a few engineers with the right platform can serve thousands. I now have a personal data point for a sharper version of that claim: one engineer, with the right agents, ships what used to take the few.