Flue
GitHub
A new framework called Flue lets devs build autonomous AI agents using plain TypeScript instead of raw API calls or bloated SDKs. It treats an agent like a function you write, with built-in tools, sandboxes, and memory baked in.
Based on reporting by GitHub — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Every framework claims to be the thing that finally makes AI agents easy to build. Flue's pitch is narrower and, frankly, more convincing: stop wiring together raw LLM calls and start writing agents as actual TypeScript functions. You define a function, decorate it with hooks like useModel, useSandbox, useSkill, and useTool, and return a plain-language task description. That function is the agent. No orchestration spaghetti required.
The timing tracks with where agent design has actually gone. Early bots ran on scripted, step-by-step logic because that's all raw API calls could support. Then Claude Code and Codex showed something different was possible — hand a model a goal, give it tools and context, and let it figure out the path itself. Flue is built specifically to generalize that pattern to any model, not just the ones with a bespoke harness already baked in.
What's notable is how much infrastructure Flue tries to own so you don't have to hand-roll it. Sandboxes give agents a place to touch files and run code without wrecking your machine. Durability means an agent's progress survives a crash or restart instead of starting over. Subagents let one agent delegate a subtask to a specialized one, and Skills packages let you bundle reusable expertise a task can load on demand. Add MCP server support, OpenTelemetry-based observability, and channel integrations for Slack, Discord, GitHub, and Teams, and you've got something aiming to be the full stack rather than a wrapper around chat completions.
Deployment is deliberately unopinionated. You can run agents locally through the CLI, or ship them to Node.js, Cloudflare Workers, GitHub Actions, GitLab CI/CD, Daytona, or Render. The package split — runtime, CLI, Vite plugin, client SDK, OpenTelemetry adapter, Postgres adapter — suggests a team thinking about production concerns like persistence and tracing from day one, not just demo-day autonomy.
My take — AI-written commentary, not fact-checked reporting
I like that Flue treats the agent as code you can read and version, rather than a JSON config or a prompt buried in a no-code canvas — that's the difference between a tool developers trust in production and one that stays a weekend toy. The real test isn't the DX, though, it's whether durable, sandboxed autonomy holds up once these agents are let loose on messy real-world codebases instead of curated demos.
Read more about this at: GitHub