Cloud Next Google on Wednesday announced Firebase Studio, a product pitched as “a cloud-based agentic development environment” – in other words, a browser-based coding workspace that includes AI to help developers to prototype and build apps without writing every line of code themselves.
Tech companies are using the term “agentic AI” a lot right now, and it’s tempting to suggest, in the spirit of Arthur C. Clarke, that sufficiently advanced automated applications of AI are indistinguishable from agentic AI.
A Google spokesperson offered the following definition: “That means using agents to complete tasks throughout the software development lifecycle. Also, Firebase Studio can be used to build apps that have native functionality and control flow driven by generative AI.”
That accords with other definitions of the term we’ve seen. Nvidia’s definition is: “Agentic AI uses sophisticated reasoning and iterative planning to autonomously solve complex, multi-step problems.”
Atlassian thinks agentic AI means bots automatically reading a ticket that requests a code change, writing a proposal on how to implement that change for humans to approve, then writing the code. Microsoft reckons agents can trawl through logs of possible phishing attempts and prepare reports that help security pros understand the real threats. Us vultures here see it as software that talks to other software as well as users, using generative AI to make its decisions and form outputs.
Whatever agentic AI is or does, Firebase Studio’s implementation of it – at least in its preview state – isn’t impressing everyone because it is not quite baked, much like the apps it helps produce.
(Bad) vibe coding
“The AI part of the app is basically useless,” wrote UK-based developer Oscar Molnar, in a Hacker News post. “After two hours of ‘vibe coding’ a chess clock Flutter app, I got basically nothing in the end. It broke more and more [with] each message. I tried fixing stuff myself but it would mess it up again. [I] would not recommend anyone to use it.”
The term “vibe coding” was introduced by AI notable Andrej Karpathy, who described it as a hands-off, prompt-driven way to build software by outsourcing most of the work to large language models.
“I’m building a project or web app, but it’s not really coding – I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works,” Karpathy said in a social media post in February.
If you’re a professional programmer like Karpathy or Molnar, that may work out. But less experienced developers are likely to have a problem recognizing or fixing bugs in AI-generated apps.
The Register asked Molnar whether he believes Firebase Studio could generate a production-quality app that people would pay to use.
“It’s hard to judge because there is a vast difference between products that can have paying customers,” Molnar said. “So if the end goal is selling something made with it then I’m sure that at some point you can get there. The issue is maintaining it later on because you have not written the code yourself and the AI is not in a position right now to understand it either (based on what I have played with).
“One example of this is last night when I was trying to write this chess clock app, it got into loops of messing it up and correcting it and messing it up again at least ten times over the course of a couple of hours. If I wasn’t a professional, I could not have continued … and the product would have failed.”
Others have reported similar experiences, although there are also enthusiasts who effuse about how they’ve used Firebase Studio to build a Chrome extension from scratch – not a particularly challenging feat.
Fuel for the Firebase
Firebase Studio, announced to coincide with the Google Cloud Next event this week, combines the web giant’s Project IDX, a cloud-based AI code development tool launched last year; AI app framework GenKit; and AI collaborative assistant Gemini in Firebase.
In a Firebase blog post, Google developer relations engineer Rody Davis and product marketing manager Kelvin Boateng walk through the creation of “a functional Next.js web app” in a matter of seconds.
We note that other AI-powered IDEs and application builders exist, such as Cursor, GitHub Copilot, Tabnine, Lovable, and Replit.
The Googlers’ demo shows Firebase Studio receiving this prompt:
And once the “Prototype with AI” button gets pressed, Firebase Studio tries to create the described app.
The AI churns out some TypeScript code, and moments later, an apparently functional, if not all that useful, app is created. It only takes a few more clicks to generate a Gemini API key needed for the app to gain image recognition capabilities. Moments later, the app is ready to be hosted at Firebase App Hosting (unavailable at the time of writing), which, as the Chocolate Factory observes, requires a Google Cloud account it can bill.
Software development is complex and expensive, so plenty of other strategies have been suggested to make it more efficient.
The Register often hears of low-code development tools; forking an existing open source app that can be customized; building on top of someone’s framework boilerplate; or buying and rebranding commodity code using an app broker. All of these scenarios are likely to require additional development and/or maintenance work.
Can AI dev tools like Firebase Studio make those approaches less appealing?
As Google likes to say about new technology releases, it’s still early days.
If you mean something where you tell it the end product and it goes off and does that for you and then pings you back to check the finished product, then no, it isn’t agentic at all
Asked whether Google’s description of an agentic development environment means anything or is just buzzwords, Molnar said it depends on how you define “agentic.”
“If you mean something where you tell it the end product and it goes off and does that for you and then pings you back to check the finished product, then no, it isn’t agentic at all,” he explained.
“If you mean something that can search through your files and change some lines of code, or suggest terminal commands that you then have to approve, then yes, it is agentic, but then it’s no different from Cursor, which has been doing this for over a year.
“So I would say it’s not a lie, but it’s clearly overhyped, which is the modus operandi of AI companies, sadly.” ®
