New UI foundations for the agentic era
AI agent UI frameworks are emerging as tools that structure both user interfaces and machine-driven workflows so that humans, large language models, and infrastructure can safely share control over applications. ArrowJS 1.0 and Vercel’s eve framework highlight two contrasting answers to this problem, one centered on reactive DOM updates and the other on durable, directory-based agents. ArrowJS positions itself as a minimal reactive UI layer that agents can understand and even generate, while eve focuses on turning each agent into a deployable unit containing prompts, tools, and workflow logic. Together they show how the agentic era of development pushes UI infrastructure beyond rendering views and towards coordinating long-running conversations, tool calls, and sandboxed code. For developers, the question is less “which framework is faster?” and more “which model fits the way humans and agents will collaborate in this app?”.
ArrowJS 1.0: reactive UI built on web primitives
ArrowJS 1.0 recasts a tiny reactive library as one of the first AI agent UI frameworks, emphasizing primitives that both developers and language models know: JavaScript modules, tagged template literals, and direct DOM access. There is no JSX, compiler, or required build step, and the core runtime is under 5kb while still benchmarking on par with Vue 3. The entire API is centered on three functions—reactive, html, and component—which keep the documentation small enough to fit in less than 5 percent of a 200k-token context window, making the library easier for agents to learn and apply. The new @arrow-js/sandbox package runs component logic inside a QuickJS WebAssembly realm, allowing untrusted agent-generated UI code to run without iframes or eval. Around the core, optional packages handle async components, server rendering, and hydration, while an npx-installed “skill” teaches coding agents how to integrate Arrow into existing projects.
Vercel eve: agents as directories and durable workflows
The Vercel eve framework takes a different path by treating each agent as a directory of files that describe its model, system prompt, and tools. One file selects the model and relies on Vercel’s AI Gateway for provider fallbacks, another holds the system prompt in Markdown, and each tool is a TypeScript file whose filename becomes the tool name, with no extra registration. Eve compiles this directory into a running agent, where every conversation is a durable workflow using Vercel’s open-source Workflow SDK, checkpointing each step so sessions can recover from crashes or pauses. Each agent receives its own sandbox for code it writes, and tools can require human approval. Built-in channels for Slack, Discord, Microsoft Teams, Telegram, Twilio, GitHub, and Linear connect agents to users, while every run emits an OpenTelemetry trace into Vercel’s observability dashboard. Eve is open-source under Apache 2.0 and currently runs on Vercel in public preview.
Reactive vs directory-based: different philosophies for agentic development
ArrowJS and the Vercel eve framework represent two philosophies for agentic era development. ArrowJS stays close to the browser, focusing on reactive state and DOM updates so agents can output HTML templates and simple JavaScript without learning a complex component model. This makes Arrow fit naturally into existing front-end stacks, but also leaves developers to design their own higher-level workflow and orchestration patterns. Eve, in contrast, treats agents as first-class deployment units: directories that bundle prompts, models, tools, sandboxes, workflows, and observability into a single artifact. According to Vercel, agents already trigger around 29 percent of deployments on its platform, and that share is expected to reach half, which helps explain the workflow-heavy design. For teams building rich interfaces, ArrowJS can power the UI while eve coordinates durable, tool-using agents behind the scenes, pointing toward layered architectures instead of one framework to rule everything.
Developer experience and what comes next
For developers, the choice between these AI agent UI frameworks comes down to where they want structure. ArrowJS offers a light reactive layer that feels close to vanilla JavaScript: template literals for markup, minimal API surface, and optional WASM sandboxing for agent-written UI logic. It trades ecosystem depth and guardrails for simplicity and agent readability, which may appeal to teams experimenting with AI-generated interfaces. Eve gives a heavier, infrastructure-oriented experience: agents are directories, conversations are workflows, and deployment is tied to the vercel deploy command, with observability and sandboxing built in. Developers start agents from the terminal and connect them to channels without custom plumbing. As more applications rely on agents not only to call APIs but to shape interfaces and long-running tasks, the likely pattern is hybrid: reactive UI frameworks like ArrowJS at the edge, and directory-based frameworks like eve coordinating the agents behind them.
