Origin

Personal note from the author. First-person and unpolished by intent — the story of a frustration that became a library.

---

The frustration

Earlier this year I was trying to build a simple thing: an AI chat interface with streaming, a couple of tools, and persistent memory. Not an agent that cures cancer — just a chat that remembers yesterday's conversation and can call a function.

I went shopping in the JavaScript ecosystem.

Vercel AI SDK got me 80% of the way in an hour. Streaming — excellent. Tool calling — clean. Then I wanted an autonomous loop, multi-agent delegation, a runtime that survives a crash. Nothing. Vercel AI SDK is a beautiful chat SDK. It is not an agent framework, and it never pretended to be.

LangChain.js had all the abstractions I wanted. It also had 200MB of dependencies, leaked abstractions at every layer, and every example on the internet was outdated within six weeks. I spent two days debugging a RunnableSequence and writing no product.

assistant-ui had beautiful components. I counted 53 before I gave up trying to figure out which one I needed.

Mastra was promising and new — exactly my problem. I loved the direction but the pieces I wanted most weren't there yet.

MCP solved tool interop brilliantly. It solved nothing else. You still had to write the UI, the runtime, the memory, the streaming.

So my weekend project became: paste LangChain code into Claude, remove half of it, add bits of Vercel AI SDK, wire up a homebrew ReAct loop, glue a Pinecone client onto the side. Three days in, I had something that worked. Four days in, I couldn't debug it. Five days in, I deleted it all.

The observation

The JavaScript ecosystem has given us a beautiful pattern, over and over again: a small, opinionated core with a radically composable ecosystem around it. React. Express. Vite. Fastify. Drizzle. Hono.

The agent space has not yet found its React. It has a monolith (LangChain), a narrow product (Vercel AI SDK), component libraries without runtimes, runtimes without UI, and a protocol (MCP) with no opinion about the application layer.

The observation is simple: we don't need another framework. We need a kit — a family of small, well-contracted packages where every piece works alone and every combination works together.

The bet

AgentsKit.js is a bet on three claims:

1. JavaScript will be the language of agent applications, because it's where the users are, where the deployment is cheap, and where the ecosystem is most plug-and-play.
2. Agents don't need a framework; they need contracts. A Tool is a function. A Memory is a store. A Runtime is a loop. Formalize the contracts, keep the packages small, let users assemble.
3. Plug-and-play beats opinionated all-in-one. If you outgrow any package, swap it. If you don't need one, don't install it. The core is 10KB and stays that way.

The promise

AgentsKit.js will never become the thing I was frustrated with. It will stay small. It will stay composable. It will stay honest — about what it does and what it doesn't. It will tell you when you should use something else.

If AgentsKit.js ever starts feeling like LangChain — bloated, magical, hard to debug — we failed. If it ever stops feeling like plain JavaScript you already know — we failed. The manifesto is how we hold ourselves accountable.

The invitation

If you've felt the frustration above, this is for you. If you've built agents in JavaScript and thought "there has to be a better way" — this is that way, or at least our attempt at it.

Open an issue. Open an RFC. Build a package. Break our assumptions. Make us better.

— Emerson Braun, April 2026