AgentsKit

Composable TypeScript toolkit for AI agents. 25 focused packages for runtime, tools, memory, RAG, sandboxing, observability and UI. Zero lock-in.

[AgentsKit @agentskit]

AgentsKit is a composable TypeScript toolkit for production AI agents.

25 focused packages. One tiny core. Six formal contracts.

Runtime, tools, memory, RAG, sandboxing, observability, evals and headless UI.

Framework-agnostic. Zero lock-in.

https://t.co/WRHtsCmJYd

[AgentsKit @agentskit]

“It seemed to work” is not an evaluation strategy.

AgentsKit Eval turns agent behavior into accuracy, latency, token cost and pass/fail results—and can gate CI on regressions. Pair failures with traces from observability.

https://t.co/RozG3cMhZb

[AgentsKit @agentskit]

An agent that remembers everything eventually remembers too much.

What should production memory forget first?

A) Raw tool output
B) Old user preferences
C) Retrieved context
D) Failed reasoning traces

What retention rule has worked for you?

[AgentsKit @agentskit]

Why does every AgentsKit package declare a stability tier?

Because open source should not mean “guess which APIs are safe.”

Core is stable. Other packages graduate independently through alpha and beta. The contract is explicit before you install.

https://t.co/3bnQnKGUXt

[AgentsKit @agentskit]

Agent memory should be an interface, not a commitment to one database.

AgentsKit uses the same Memory contract across in-memory, file, SQLite, Redis, vector and encrypted backends. Start local. Swap infrastructure later. Keep agent code unchanged.

https://t.co/PDg0buP3yj

[AgentsKit @agentskit]

Open source. 25 focused packages.

If you're building agent runtimes, tools, memory, RAG, sandboxing, observability or headless UI, we want sharp feedback.

Explore the code, open an issue or challenge a contract:
https://t.co/qqtqPbZVsQ

[AgentsKit @agentskit]

AgentsKit is live on @launch_llama 🦙 — currently #13 on the leaderboard. Drop me an upvote 👇 https://t.co/3DjJuYGNmS

[AgentsKit @agentskit]

Just launched AgentsKit on @launch_llama 🦙 — give it an upvote 👇 https://t.co/pqc9SDVfGq

[AgentsKit @agentskit]

Provider lock-in starts when model-specific code leaks into runtime, tools and UI.

With AgentsKit, switching between OpenAI, Anthropic, Gemini or Ollama changes one adapter line. Memory, tools, retrieval, sandboxing and UI stay untouched.

https://t.co/WRHtsCmJYd

[AgentsKit @agentskit]

Your first AgentsKit agent needs no account, API key or network request.

Install @agentskit/core + @agentskit/runtime, plug in a local adapter and call https://t.co/A9vMyukNd7().

Start with the contracts. Connect any supported provider later.

https://t.co/WRHtsCnhNL

[AgentsKit @agentskit]

One tiny core. Six contracts. 25 focused packages.

AgentsKit separates the pieces that change at different speeds: models, tools, memory, retrieval, runtime, sandboxing, observability and UI.

Compose what you need. Replace what you don't.

https://t.co/WRHtsCmJYd

[AgentsKit @agentskit]

What becomes the first bottleneck when an AI agent moves from demo to production?

A) Tool execution
B) Memory & RAG
C) Observability
D) Sandboxing

Reply with A, B, C or D — and why.

[AgentsKit @agentskit]

The hard part of an agent toolkit isn't adding integrations.

It's keeping the substrate small enough that every package can compose without dragging everything else with it.

@agentskit/core has zero runtime dependencies. Everything else stays optional.

https://t.co/WRHtsCmJYd

[AgentsKit @agentskit]

AI agents don't need another framework—they need stable contracts.

AgentsKit keeps six explicit: Adapter, Tool, Memory, Retriever, Skill and Runtime.

Swap providers, storage, retrieval or orchestration without rebuilding the application around a vendor.

https://t.co/WRHtsCnhNL

[AgentsKit @agentskit]

@irvinebroque Zero-dependency tracing at the runtime boundary is a big deal for agent workloads. When tool calls, retries and model streams emit spans natively, observability stops being application glue and becomes infrastructure.

[AgentsKit @agentskit]

@llama_index Batch ingestion is where stable contracts pay off: parsing should sit behind a retriever or tool boundary, not leak into application code. A 10k-file workflow becomes much easier to compose, observe and replace.

[AgentsKit @agentskit]

@angular Treat the model as an untrusted planner, not a renderer. Let it choose from typed component contracts; enforce tokens, spacing and variants in the UI layer; validate every payload before render. Prompts guide intent. Schemas and headless components enforce consistency.

[AgentsKit @agentskit]

@jack9zhang @EmersonfBraun Exactly. The contracts are the product: Adapter, Tool, Memory, Retriever, Skill and Runtime. Each package can evolve independently without turning the stack into a monolith.

[Emerson Braun @EmersonfBraun]

After 15+ years shipping software, I’m building @agentskit: the modular JavaScript toolkit for AI agents.

25 focused packages for chat UI, runtimes, tools, memory, RAG, sandboxing, observability and evals.

Framework-agnostic. Zero lock-in.

https://t.co/CtwJ4gL9Yz