I saw this and thought, if this doesn't get me to give it a go, nothing will.

Less than 45 minutes after signing up for fly.io, I have a multi-room tic tac toe game deployed.

https://tic-tac-toe-cyber.fly.dev/

I had it build the game, opting for a single room at first to see if that worked. Then I had it add multiple rooms on a different git branch in case that didn't work. It worked great.

I learned very little about elixir, phoenix, or deploying to fly.io up to this point, and I already have a nice looking app deployed and running.

I know a lot of devs will hate that this is possible, and it is up to me now to look at the steps it took to create this and really understand what is happening, which are broken down extremely simply for me...

I will do this because I want to learn. I bet a lot of people won't bother to do that. But those people never would have had apps in the first place and now they can. If they are creating fun experiences and not banking apps, I think that is still great.

You guys have been releasing amazing things for years only to be poorly replicated in other languages years later.. but you really outdid yourselves here.

I'm blown away.

edit: is there a way to see how much of my credits were used by building this?

1. Remote agent - it's a containerized environment where the agent can run loose and do whatever - it doesn't need approval for user tasks because it's in an isolated environment (though it could still accidentally do destructive actions like edit git history). I think this alone is a separate service that needs to be productionized. When I run claude code in my terminal, automatically spin up the agent in an isolated environment (locally or remotely) and have it go wild. Easy to run things in parallel

2. Deep integration with fly. Everyone will be trying to embed AI deep into their product. Instead of having to talk to chatgpt and copy paste output, I should be able to directly interact with whatever product I'm using and interact with my data in the product using tools. In this case, it's deploying my web app

I worked all day on a Phoenix app we’re developing for ag irrigation analysis. Of late, my “let’s see what $20/mo gets you” is Zed with its genetic offerings.

It actually writes very little Elixir code for me. Sometimes, I let it have a go, but mostly I end up rewriting that stuff. Elixir is fun, and using the programming model as intended is enlightening.

What I do direct it to write a lot is a huge amount of the HEEX stuff for me. With an eventual pass over and clean it up for me. I have not memorized all of the nuances of CSS and html. I do not want to. And writing it has got to be the worst syntactic experience in the history of programming. It’s like someone said people lisp was cool; rather than just gobs of nested parentheses, let’s double, nay triple, no quadruple down on that. We’ll bracket all our statements/elements with a PAIR of hard to type characters, and for funsies, we’ll make them out different words in there. And then when it came to years of how to express lists of things, it’s like someone said “gimme a little bit of ini, case insensitivity, etc”. And every year, we’ll publish new spec of new stuff that preserves the old while adding the new. I digress…

I view agentic coding as an indictment on how bad programming has gotten. I’m not saying there wouldn’t be value, but a huge amount of the appeal, is that web tech is like legalese filled with what are probably hidden bugs that are swallowed by browsers in a variety of u predictable ways. What a surprise that we’ve given up and decided the key tools do the probabilistic right thing. It’s not like we had a better chance of being any more correctly precise on our own anyway.

As an Elixir enthusiast I've been worried that Elixir would fall behind because the LLMs don't write it as well as they write bigger languages like Python/JS. So I'm really glad to see such active effort to rectify this problem.

We're in safe hands.

I have a question about how you manage context, and what model you use. Gemini seems the best at working with give context windows right now, but even that has its limitations. Thinking about working with Claude Code, a fair bit of my strategizing is in breaking down work and managing project state to keep context size manageable.

I'm watching the linked video and it's amazing seeing it in action, but I'm imagining continuing to work on a project and wondering if it will start losing its way so to speak. Can you have it summarize stuff, and can you start a session clean with those summaries, and have it "forget" files it won't need to use for this next feature, etc?

I've been daydreaming of an agentic framework that maximally exploits BEAM. This isn't that, but maybe jido[0] is what I'm looking for.

0. https://github.com/agentjido/jido

hard to put confidence in AI vibe hacks when the basic stuff just doesn't work.

* (Mix) The database for Myapp.Repo couldn't be created: killed

Few issues:

1. The 150 message limit is understable but it suddenly pops up and you lose significant work. I was working on UI mockup and just as I had finished and was ready to go on implementation, this limit appeared and significant part of my work was lost. 2. After the first credit, the credit seems to exhaust pretty fast which makes it expensive, especially when you are trying it out. 3. Also I don't understand when you ask it to prototype different screens, why does it overwrites the same file. 4. It is not able to stop to seek user feedback but keeps trying different approach which kind of exhausts the credit. It would be nice if it describes its approach, so the human developer can provide their feedback. 5. It seems it is using OpenAI because it is often self-congratulatory to the point of being annoying sometimes.

This is a tangential comment and should not detract from what Chris and team have created. I think closing the loop between agent and the running output is a great/critical step forward.

However, I find using AI to build transitional Apps with a UI is a bit like improving the way automobile steering wheels are made. In a world that soon won't need steering wheels at all.

If the AI is so good to write the code for an App, how much longer before you won't need those Apps in the first place? And then the question is, what will fill the role that Apps play today.

Feels like we're getting into a weird situation if LLM providers are publishing open source agentic coding tools and OSS web app frameworks are publishing closed source/non-BYOK agentic coding tool. I realize this may not be an official "Phoenix" project but it seems analogous to DHH releasing a closed-source/hosted "Rails.new" service.

I find Claude to have quite a bit of problems trying to navigate changesets + forms + streams in my codebase, just wondered if you had any tips of making it understand better :)

And thinking about it made me realize that soon there will be a completely different programming language used solely by coding agents. ChatGPT gives an interesting take on this, "The fundamental shift is that such a language wouldn’t be written or read, but reasoned about and generated. It would be more like an interlingua between symbolic goals and executable semantics, verbose, unambiguous, self-modifying, auto-verifiable, evolving alongside the agents that use it").

I've been working with Phoenix a lot the last few months, and I like it a lot. But I do get the sense that the project suffers from wanting to perpetually chase the next new thing, even when that comes at the expense of the functional elegance and conceptual cohesiveness that I think is Phoenix' main strength.

LiveView is a great example. It's a very neat bit of tech, but it's shoe-horned surprisingly awkwardly into Phoenix. There's now a live view and non-live view way to do almost everything in Phoenix, and each has their own different foibles and edge cases. A lot of code needs to work with both (e.g. auth needs to happen at both levels, basically), meaning a surprising amount of code needs to have two, nearly identical variants: one with traditional Plug idioms, and then another using LiveView equivalents. Quick little view helpers end up with either convoluted 'what mode am I in?' branching, or (more likely) in view-mode-dependent wrappers around view-mode-independent abstractions. This touches even the simplest helpers (what is the current path?) and becomes more cumbersome from there. (And given the lack of static analysis for views, it can be non-trivial to even find out what is and isn't actually working where.)

Not every website should be a live view (e.g. hiking directions, for example), but that is clearly the direction of travel in Phoenix. Non-live views get the disparaging moniker 'dead views', and the old Phoenix.HTML helpers have been depreciated in favour of <.form />-style live components. The generators depend on those, plus Tailwind, Hero Icons and (soon) DaisyUI, all fetched live from various places on the Internet on build. This tight coupling to trendy dependencies will age poorly, and it makes for bumpy on-boarding (opinionated and tightly coupled isn't necessarily a smoother experience, just a more inflexible one).

So with all of that in mind, while I'm not shocked to see Phoenix jump on the vibe coding hype train, I guess I am disappointed.

The revelation that AI is now writing PRs for Phoenix itself is not confidence inspiring. I rely on frameworks like Phoenix because I don't want to have to think about the core abstractions around getting a website to my users; I want to focus on my business logic. But implicit in that choice is the assumption that someone is thinking about those things. If it's AI pushing out Phoenix updates now, my trust level and willingness to rely on this code drops dramatically. I also do not expect Phoenix' fraying conceptual cohesiveness to get any better if that's the way we're headed.

Phoenix is still an amazing piece of tech, but I wish I felt more at ease about its future trajectory.

Having a full stack that is easy to use as a learning sandbox is incredibly helpful in that regard, so this looks amazing.

I couldn't get Tidewave working but I must try again to see if Tidewave with Claude Code would offer this level of awesome.

ps. @fly - please let me buy more credit, I just get an error!

Just a clarifying question since I'm confused by the branding use of "Phoenix.new" (since I associate "Phoenix" as a web framework for Elixir apps but this seems to be a lot more than that).

- Is "Phoenix.new" an IDE?

- Is "Phoenix.new" ... AI to help you create an app using the Phoenix web framework for Elixir?

- Does "Phoenix.new" require the app to be hosted/deployed on Fly.io? If that's the case, maybe a naming like "phoenix.flyio.new" would be better and extensible for any type of service Fly.io helps in deployment - Phoenix/Elixir being one)

- Is it all 3 above?

And how does this compare to Tidewave.ai (created as presumably you know, by Elixir creator)

Apologies if I'm possibility conflating topics here.

Is it possible to get that headless Chrome browser + agent working locally? With something like Cursor?

- ability to run locally somehow. I have my own IDE, tools etc. Browser IDEs are definitely not something I use willingly.

- ability to get all code, and deploy it myself, anywhere

---

Edit: forgot to add. I like that every video in Elixir/Phoenix space is the spiritual successor to "15-minute rails blog" from 20 year ago. No marketing bullshit, just people actually using the stuff they build.

These other details that are not “just coding” are always the biggest actual impediments to “showing your work”. Thanks for making this!! Somehow I am only just discovering it (toddler kid robbing my “learning tech by osmosis” time… a phenomenon I believe you are also currently familiar with, lol)

The LLM chat taps out but I can't find a remaining balance on the fly.io dashboard to gauge how I'm using it. I _can_ see a total value of purchased top ups, but I'm not clear how much credit was included in the subscription.

It's very addictive (because it is awesome!) but I've topped up a couple of times now on a small project. The amount of work I can get out the agent per top-up does seem to be diminishing quite quickly, presumably as the context size increases.

PS: Why can't I get IEx to have working command-line history and editing? ;-P

2. How do you handle 3rd party libraries? Can the agent access library docs somehow? Considering that Elixir is less popular than more mainstream languages, and hence has less training data available, this seems like an important problem to solve.

I was curious what the pricing for this is? Is it normal fly pricing for an instance, and is there any AI cost or environment cost?

And can it do multiple projects on different domains?

How do you protect the host Elixir app from the agent shell, runtime, etc

From time to time it tries to do something a little old-school, but nothing significant really. It's very capable at spitting out entire new features, even in liveview.

Over all the experience has very productive, and at least on-par with my recent work on similar sized python and nextjs applications.

I think because I'm using mostly common and well understood packages it has a big leg up. I also made sure to initialise the phoenix project myself to start with so it didn't try to go off on some weird direction.

If Phoenix.new helps solve that problem, I’m all for the effort. But otherwise, the sole focus of the community leaders of Elixir should be squarely and exactly focused on creating the incentives and dynamics to grow the base.

Compare, for example, Mastra in TypeScript or PydanticAI in Python. Elixir? Nothing.

Not here to bash. It’s more just a disappointment because otherwise I think nothing comes close.

If you want to take your website and business down, use ChatGPT-4o's code

My experience with software development is maybe different than yours. There's a massive amount of not-yet-built software that can improve peoples' lives, even in teeny tiny ways. Like 99.999% of what should exist, doesn't.

Building things faster with LLMs makes me more capable. It (so far) has not taken work away from the people I work with. It has made them more capable. We can all build better tools, and faster than we did 12 months ago.

Automation is disruptive to peoples' lives. I get that. It decreases the value of some hard earned skills. Developer automation, in my life at least, has also increased the value of other peoples' skills. I don't believe it's anti worker to build more tools for builders.

The days are numbered where humans are sitting typing out code themselves.

It's akin to the numbered days of type writer secretaries of the 20th century.

Overall I think we would all be happier if efficient machines take away the drudgery of our daily work and allow us to focus on things that really matter to us. . . as long as our basic needs are met.