Replit’s iOS app hasn’t received an update since January. Its ranking in Apple’s free developer tools category has slid from first place downward. The reason traces back to a decision made in Cupertino. Apple invoked a years-old rule to block changes to apps that let users generate and preview new software on the fly.
The standoff highlights a larger tension. Traditional distribution systems expect software to remain fixed after release. New AI systems produce code that shifts with every prompt. One side reviews binaries. The other composes behavior at runtime. The gap between them grows wider by the month.
The Information first detailed Apple’s actions in March. Review Guideline 2.5.2 requires apps to stay self-contained. They cannot download or execute code that adds features after approval. Replit and similar tools ran afoul by embedding generated previews inside their own interface. Apple suggested a workaround. Route previews through Safari instead. Yet enforcement tightened. An app called Anything lost its listing entirely despite multiple attempts to comply.
From Apple’s view the logic holds. Reviewers inspect one artifact. When that artifact spawns unlimited variants based on user input, the inspection loses meaning. The wrapper passes scrutiny. The code it summons does not. This isn’t new ground for the App Store. Past battles involved interpreters, JavaScript execution, and downloadable content. Each time the platform owner redrew the boundary. The core demand stayed constant. Software must hold still long enough for inspection.
But software no longer holds still. Adaptive systems generate fresh implementations per session. Two users running the same product can experience different features. A reported bug may vanish before support teams reproduce it. Version numbers lose their anchor when the binary mutates through model output. Documentation screenshots fail to match reality. The entire apparatus of release notes, bug trackers, and package registries rests on an assumption now under pressure.
Iris laid out this argument in Adaptive Software on May 5. “The reviewable artifact and the running artifact are not the same kind of thing,” the author wrote. The post frames Apple’s conflict with Replit as the first major collision. Other layers of the distribution stack will face similar questions soon.
OpenAI took the opposite path. In October 2025 the company unveiled its Apps SDK at DevDay. It built on the Model Context Protocol, an open standard first shaped at Anthropic. By early 2026 a directory inside ChatGPT hosted integrations from Spotify, Zillow, Canva, Stripe, and even Replit. These aren’t binaries. They consist of MCP servers paired with UI components the model renders inline during conversation.
Discovery changes. Users don’t browse and select. The model surfaces capabilities based on context. Mention buying a house and Zillow appears without navigation. The app gains no fixed interface. It offers components assembled turn by turn. OpenAI maintains a review process and quality standards. Payments flow through an Agentic Commerce Protocol in beta with Stripe. The gatekeeper model persists, yet the unit distributed has evolved.
Recent coverage shows the friction intensifying. Futurism reported two days ago that vibe coding startups face a meltdown over Apple’s blocks. The Financial Times detailed complaints from developers who see the rules as outdated. Replit told the FT it felt surprised and disappointed after years of approved updates with the same capabilities.
At TechCrunch’s StrictlyVC event on May 1, Replit CEO Amjad Masad didn’t mince words. He called Apple’s stated reasoning “a lie” and signaled willingness to prove it in court. TechCrunch captured the exchange. Masad noted Replit’s net revenue retention climbing toward 300 percent even as mobile updates stalled. The company continues to push despite the hold.
Productivity data adds another layer. Engineering platforms report high code acceptance rates from AI tools, often 80 to 90 percent. Yet downstream churn tells a different story. TechCrunch examined the phenomenon in April. GitClear’s January report found regular AI users generated 9.4 times higher code churn than non-AI counterparts. Faros AI measured an 861 percent rise in churn under heavy adoption. Jellyfish tracked 7,548 engineers and saw doubled throughput at ten times the token cost. Volume arrives. Sustainable value proves harder to pin down.
These metrics expose limits in current agent setups. Many tools function as sophisticated wrappers around models like Claude or GPT. They accelerate prototypes but struggle with production demands such as security, scalability, and long-term maintenance. A Medium post from last week described one founder generating $121,000 in 72 hours with a vertical SaaS wrapper built via vibe coding. Success stories exist. So do warnings about demo-to-production gaps.
Platforms race to claim the next distribution surface. Apple Intelligence positions Siri as agent layer for its hardware. Gemini does similar work on Android. Microsoft embeds Copilot across Windows and Office. Browser-based agents from OpenAI, Perplexity, and Anthropic bet that intent-driven interfaces inside Chrome will capture user actions. The unit of distribution climbs. Binaries give way to capabilities, then to constraint sets that fulfill expressed needs.
The contradiction sits on the same devices. ChatGPT passed Apple’s review. It now hosts its own directory that summons third-party functionality in ways that skirt the spirit of the rules applied to Replit. As that directory scales, Apple must choose. Enforce the guideline against one of the world’s most downloaded apps and invite antitrust heat. Or carve out exceptions for model-mediated runtimes. Either path redraws power balances across mobile platforms.
No public memo has appeared. The decision will likely emerge through a specific review ruling in Cupertino. Reviewers once examined code. They now confront artifacts that dissolve into runtime behavior. The old premise bought decades of stability. New tools make that stability optional.
Developers adapt in real time. Terminal agents like Claude Code handle complex codebases with deep context and sub-agent orchestration. IDE extensions add guardrails and workflow controls. Teams encode conventions in files such as AGENTS.md to reduce repeated instructions. Yet the core challenge remains. How do organizations maintain ownership when generation outpaces oversight? How do they measure output when accepted code creates hidden debt?
The fight isn’t solely between Apple and Replit. It pits two views of software against each other. One treats it as a fixed artifact suitable for inspection and versioning. The other sees it as fluid, responsive to context, generated fresh for each task. The first built the infrastructure that powered the last forty years. The second leverages models that didn’t exist until recently.
Platforms that accommodate motion stand to gain. Those anchored to stillness risk fading into memory alongside earlier formats that solved real problems inside assumptions that didn’t endure. The memo from Cupertino will arrive. The only variable is how much time passes before it accepts the change already underway.
WebProNews is an iEntry Publication