Agent frameworks are becoming what web frameworks were a decade ago: the default way you start building products.

An AI agent is a program, that hands a task to a LLM and lets it work: the model reads what you want, picks a tool to call, runs it, looks at the result, and repeats until the job is done. An agent framework is the scaffolding around that loop, so you don't rebuild the same plumbing for every project.

A framework usually handles:

Mastra is one of the leading AI agent frameworks in TypeScript, with tons of useful videos and documentation.

Mastra gives lots of powerful primitives for building agents: workflows with branching, durability and retries, memory and RAG so the agent can pull in your own data, traces and evals to see what it did.

Mastra comes with local Studio, the UI for testing agents while you build them.

Eve is an agent framework from Vercel. With Eve there is no code to wire the agent together. The agent is a directory, and each file has a fixed job: the prompt, the tools, the schedules. You add a capability by adding a file, and Eve turns the folder into a running agent.

That structure keeps a project easy to read: you can open the folder and see everything the agent knows and can do.

Eve runs on Vercel primitives, so it comes with powerful built-in features: durable execution, sandboxed VMs for running code, human approval steps, and evals. The catch is that Vercel can become a lock-in.

Flue takes the harness that powers a coding agent like Claude Code and makes it fully programmable. It's built on Pi, the minimal agent harness OpenClaw also runs on, and wraps skills, tools, and sandboxes around it out of the box. You describe what the agent knows (its model, skills, and instructions) instead of scripting the loop yourself.

The difference from Claude Code is that Flue doesn't assume a human is driving. You can build interactive agents that ask for input, or workflows that run start to finish with no human at all, which suits fixed, repeatable processes. Agents live in an agents/ directory, workflows in workflows/, and both pull in reusable skills.

The clever part is the sandbox. Every agent gets its own by default, but instead of booting a real Linux container, Flue runs a bash shell reimplemented in TypeScript, in memory. The model still gets grep, glob, read, and a shell, without paying container startup on every run, so you can run thousands of agents for almost nothing and opt into a real container only when a task needs it.

Deployment picks one of two targets. Node serves the agent over HTTP on Hono and runs anywhere Node does, including a VM or GitHub Actions. Cloudflare deploys it to a Worker where each agent instance maps to its own Durable Object for persistence. Channels like Slack, GitHub, and Discord are one flue add channel away, and Durable Streams log every step so a crashed run resumes where it stopped.

See also: Vercel AI SDK for lower-level streaming and tool calling without the full framework.