Shopify's mobile engineering team hit a ceiling that had nothing to do with model quality. Three or four AI agent workflows running in parallel were enough to max out a developer's laptop, with compilers, simulators, and emulators fighting for resources1. The fix was not a better prompt or a smarter model. It was infrastructure: a fleet of remote Macs with pre-built caches, SSH access for agents, attached Android emulators, and live-streamed simulator output piped to a browser. The system, called Remote Dev, is the clearest example yet of a structural shift in how coding agents are being deployed. They are no longer bolted onto a developer's existing toolchain; they are absorbing the runtime itself, from build environments to live preview to deployment. The important caveat: Shopify's 6%-to-61% PR jump cannot be cleanly attributed to infrastructure alone — model upgrades, expanded task scope, and organizational mandate all likely contributed. But the engineering team's own account centers the infrastructure investment, not model capability, as the unlock. River ran on the same underlying models before and after Remote Dev; what changed was the agent's access to build, test, and preview surfaces. That framing — infrastructure as binding constraint — is reinforced by the pattern across this month's releases.
Model capability is obviously still advancing. On ARC-AGI-2, Claude Haiku 5.5 scored 61.25 compared with Claude Haiku 4.5's 4.03. Anthropic's IPO at a $100B valuation alongside $63B in debt financing signals continued investment in raw model power. But the evidence across practitioner releases points to a different bottleneck. Amazon Web Services shipped Session Endpoints for Bedrock AgentCore Runtime, giving each agent session an isolated microVM with an authenticated URL so a developer can watch a Vite website update in real time as the agent edits files3. Google published A2UI, a framework-agnostic standard for generative UI, paired with CopilotKit's agent-integration stack, so agents can render interactive components inside a conversation rather than returning walls of text4. Warp built cloud agent orchestration into its terminal through a system called Oz, letting developers coordinate local and cloud agents from one environment6. Cognitum One released Ruflo, an open-source orchestration layer for Claude Code and Codex that turns a single assistant into a coordinated team of more than 100 specialized agents handling planning, building, testing, and security review in parallel7. Ruflo claims 12.5 million npm downloads. Each of these moves extends the agent's surface area beyond code generation into the surrounding runtime. A necessary qualification: these are vendor announcements, not adoption data. No deployment metrics, customer counts, or usage figures have been published for AWS Session Endpoints, A2UI, Warp's Oz, or Ruflo beyond that npm figure. The structural shift is visible in what vendors are building; whether practitioners are reorganizing around it at scale remains unproven.
Shopify's engineering post is the most detailed case study of what this looks like in practice. Before Remote Dev, the company's AI agent, River, ran on Linux and could edit Swift, but building and testing the iOS app still required a developer's Mac. The team solved this by setting up remote Macs with development environments and iOS simulators, connecting Android emulators through a separate Linux VM with nested virtualization because macOS does not support it natively. Agents connect over SSH, and a CLI encapsulates the complexity of the connection flags. Initial app compilation and main checkout took over 30 minutes, so the team switched to keeping a pool of warm Macs with main checked out, tools installed, apps built, and caches ready. The time from requesting a Mac to inspecting code changes on a running device dropped to a few minutes.
River's share of merged mobile pull requests grew from 6% to 61%. The system now handles rebasing branches, resolving merge conflicts, checking CI failures, addressing PR feedback, triaging incoming crashes, and reviewing code. Some automated workflows run overnight, picking up bugs, implementing fixes, testing them, and creating PRs without human intervention. The team built Live Preview to stream the iOS Simulator and Android Emulator from River's environment to a website over WebSockets, making sessions shareable so multiple people can watch and collaborate on the same running app. River can even generate explainer videos from recordings, showing the app running inside 3D phone renders with animations and narration.
ANALYSIS The individual components — remote Macs, warm caches, emulator farms — are conventional CI/CD practices that predate agents. What is new is their recomposition around an agent as the primary consumer: the warm pool exists because the agent needs sub-minute startup, the live-stream pipeline exists because the agent cannot tap a human on the shoulder, and the SSH fleet exists because the agent needs programmatic access to build surfaces that were designed for interactive use.
AWS's approach reinforces the pattern from a different angle. AgentCore Runtime runs each agent session in its own isolated microVM, and Session Endpoints forward HTTP methods, paths, query strings, request bodies, and WebSocket connections to specified ports. The design intent is to eliminate the deployment step from the review loop: instead of copying an app out and deploying it elsewhere, a reviewer opens the app where the agent built it, asks for a change, and watches the result appear in seconds.
The enterprise governance question is real. JPMorgan Chase's published framework for scaling agentic coding assistants in a regulated environment emphasizes that agentic tools differ from autocomplete because they act: they read files, run commands, and call external systems, making permissions the primary control surface rather than prompts5. The bank's model relies on centrally managed deny rules, deterministic hooks for policy, reviewed skills for repeatable workflows, and an allowlist of approved tool integrations. Published evidence on AI coding productivity is described as genuinely mixed, ranging from a 55.8% speed-up on a greenfield task to a 19% slowdown for experienced developers in mature codebases. That variance by task type and developer experience matters: it is the signature one would expect if model capability and codebase complexity were still active constraints alongside infrastructure. The thesis here is not that infrastructure is the only constraint, but that it has become the binding one for organizations that have already cleared the model-capability bar — Shopify's River was already capable of writing correct Swift before Remote Dev existed; it just had nowhere to build and test it. Developers who use assistants tend to produce less secure code, which they are also more confident in.
That security finding is worth holding against Shopify's 61% merged-PR figure, especially as Claude Code's run-rate reaches $2.5B as of February 2026 and agent-generated code volumes scale. Shopify addresses it partly through its feedback infrastructure: a classifier model called Jev observes all River conversations on mobile, tags them, and categorizes each conversation based on outcome quality, feeding failures back as input for improvement. The company runs the process iteratively so River can improve itself for future usage. But JPMorgan's framework insists that AI-generated code must go through the exact same review, testing, and security processes as human-written code. That raises the question of whether agent-dominated PR volumes can sustain the same review rigor at scale — the control plane is being built, but the review capacity to match agent throughput is not.
Three things to watch. First, whether Shopify's 61% merged-PR share for River continues to climb or plateaus, and whether the company publishes data on defect rates in agent-generated versus human-generated code. Second, whether AWS AgentCore Runtime's session-endpoint model gets adopted by other cloud providers, establishing isolated microVMs as the standard agent execution environment. Third, whether JPMorgan or another regulated institution publishes post-deployment security metrics for agentic coding at scale, testing whether the productivity-security tension identified in their framework resolves or worsens as agent PR volumes grow.