Astro Creator Fred Schott Ships Flue 2, Bringing React-Style Hooks to AI Agent Development

Fred Schott, the developer behind the Astro web framework, has released Flue 2, the first stable version of his AI agent framework. The headline addition is a system of React-inspired "Agent Hooks" that let developers compose agent behavior dynamically rather than configuring it upfront.

In Flue, an agent is represented as a JavaScript function that re-renders on every turn — meaning before each model call. Flue 2 ships with 16 built-in hooks, including useSkill(), useTool(), and useSubagent(), authored in TypeScript. Developers can also write custom hooks. According to Flue's launch post, these hooks let agents "manage their own state, listen to agent lifecycle events, and even attach different resources and capabilities dynamically to enhance themselves at runtime."

Schott, whose company was acquired by Cloudflare in January, said the hooks approach emerged after he abandoned an earlier plan to port over web-framework concepts like file-based routing. "We kind of naively ported that over to Flue, thinking — great, well, I'll put your five agents in these five files," he told Latent Space. But early users, particularly larger customers, treated their entire company as a single agent with no need for routing. That pushed him toward React's composability model instead.

Flue's core architectural claim is that every agent requires a "harness" — an environment providing the context and capabilities needed to complete tasks, as opposed to a script directing a model turn by turn. "There is no agent without a harness," Schott said. Flue is built on top of Pi, an open source minimal harness, with Flue itself acting as an opinionated layer on top — a relationship Schott compares to Vite's role beneath Astro. Hosted agents in Flue 2 are now built using Vite.

The project originated inside the Astro repository earlier this year as an issue-triage script before evolving into what Schott described in the Flue 1 launch post (early May) as being "like Claude Code, but 100% headless and programmable." Schott confirmed that many developers currently onboard onto Flue by using Claude Code itself to walk through setup, since the framework's documentation is markdown-based.

The closest competitor, according to Schott, is Vercel's eve, which also launched this year and treats the harness as a built-in feature rather than an add-on. He drew a distinction with what he called "OG agent frameworks" — Vercel's AI SDK, Cloudflare's Agents SDK, and Mastra — which are retrofitting harness support rather than building around it from the start. Schott declined to position Flue against emerging "meta-harness" projects like Databricks' Omnigent or Exo, arguing that a single API spanning multiple harnesses would blur Flue's specific definitions of skills and subagents.

Despite working at Cloudflare, Schott emphasized that Flue remains "an open source framework for every host" and said host portability is a defining principle of the project, distinguishing it from tighter platform integrations.

What this means

Flue 2 is a framework release, not a model release — no new weights, no benchmark scores, no pricing tied to inference. Its significance is architectural: it's an early attempt to establish a standard abstraction layer for agent composition, at a moment when, as Sierra CEO Bret Taylor put it, the agent ecosystem is still "in the jQuery era... not the react era." Whether hooks-style composability becomes the dominant pattern for agent frameworks — as React did for UI — depends on adoption by teams building production support and triage agents, the exact use case Schott says drove the redesign. With Vercel's eve pursuing a similar harness-first philosophy, the next competitive question is less about hooks and more about which framework's harness model developers standardize on.