Apple just put AI agents inside Xcode. Here's how we're rethinking a product built for that gap
Last week we asked whether Phoenix.vu had a product problem or a marketing problem. The honest answer might be neither. The ground moved under us.
Last week we posted here about being stuck. We'd built Phoenix.vu, an AI coding agent for Xcode, into the gap Alex Sidebar left when its team joined OpenAI. Signups were low, almost nobody came back, and nobody had paid.
That question deserves a proper answer, because we'd underweighted it.
What changed
In February, Xcode 26.3 shipped with two coding agents built in: Anthropic's Claude Agent and OpenAI's Codex. They can explore a project, update settings, build, and fix errors in a loop. Xcode also opened its tools to outside agents through MCP, so Claude Code or the Codex CLI can build your project too.
Then at WWDC, Apple showed Xcode 27 agents that plan before they code, render SwiftUI previews to check their own work, write and run tests, take a sketch as input, and split big jobs across sub-agents.
Read our own homepage pitch next to that list: "an AI agent for Xcode that writes Swift, runs builds and fixes errors." A large part of that is now just what Xcode does.
What we got wrong?
We positioned Phoenix against Alex's absence. We should have positioned it against Apple's presence.
When Alex shut down, "an AI agent that works with Xcode" was scarce, so it sounded like enough of a pitch. By the time we launched, it was a default. A developer seeing our landing page had no reason to think "I need this" rather than "doesn't Xcode already do that?" and we never answered the question.
What we still think is true
We're not pretending the built-in agents are bad. They're strong, and for some developers they're the right choice. If you already pay for Claude or ChatGPT and most of your work is SwiftUI, start there. We'll say that on our own site.
But there are real differences, and they matter to a specific kind of developer:
No subscription. Xcode's agents run on your Claude or OpenAI plan. Phoenix is pay-per-task: $1 buys 100 credits, and credits never expire. If your AI use comes in bursts, a quiet month costs nothing.
Diff-first by default. Every change is a diff you approve or reject, and Phoenix snapshots each file before writing to it, so undo doesn't depend on Git.
Safety without setup. Secret redaction and sanitized build logs are on by default.
Several models, one balance. No separate accounts or limits to track.
Whether those differences are enough is exactly what we don't know yet.
What we're changing
Answering the question head-on. We've written a straight comparison of Xcode's built-in agents and Phoenix, including when you should use Apple's instead of ours. It goes live on our blog tomorrow.
What we'd love to hear
For iOS and macOS developers:
Have you turned on Xcode's built-in agents? What did you think?
Is there anything that would make you add a second AI tool alongside them, or does built-in settle it?
Does "no subscription" matter to you, or do you already pay for an AI plan anyway?
For founders:
Has a platform ever shipped your core feature for free? Did you narrow your focus, go broader, or change who you were selling to?
How long did you give the new positioning before judging whether it worked?
The comparison goes live on our blog tomorrow, Oct 1. We'll add the link in the comments once it's up.
Built in agents raise the bar, but the model choice still shows up in edge cases like debugging loops, tool use, and how safely changes are reviewed. I put together a ByteForward video comparing GPT 6.1 Sol and Sonnet 5.5 with that practical lens.
https://youtu.be/g-Xn3M6pwoQ
Agreed that the model still matters, especially in debugging loops. That's where we see the biggest difference between a tool that finds the root cause and one that keeps patching symptoms. The review side is the part we care about most: an agent can make a build go green while quietly hiding the real problem, so we show every change as a diff before it lands. Thanks for sharing the comparison, I'll take a look.
This is the scary part of building on top of a big platform. The moment the native product becomes “good enough”, your whole positioning can disappear overnight. It hasn’t happened to me yet, but I think the key is to stay focused and not panic.
Big platforms won’t solve every use case for everyone. There’s still room for smaller products that go deeper on a specific problem or audience.
Thank you, this is exactly how it felt. The scary part wasn't that Apple built something good. It was realising our pitch only made sense while nobody else was doing it. I agree there's still room, but only if we can clearly say who we're for and what we do better, rather than just "AI in Xcode." For us that's developers who want to pay per task instead of another subscription, and who want to approve every change. We're testing whether that's enough. Appreciate the reminder not to panic.