A production-grade framework for building AI agents.
Set an OPENAI_API_KEY environment variable (get your API key here):
export OPENAI_API_KEY=<YOUR_API_KEY>
Run a simple CLI example using ts-node:
npm install
npx ts-node --experimental-specifier-resolution=node --esm example/cli.ts
By default this uses gpt-3.5-turbo; this is easy to change in example/agent.ts.
Poking around this example agent which should give you a feel for how the framework works. In particular, our goals were to create a framework that is:
- strongly typed
- fully asychronous
- compact but robust in terms of code
- easy to prototype with, particularly when dealing with complex agent patterns
- comprehensive, for example in its handling of hooks
Creating an agent requires implementing four methods and an optional regex, per the Agent interface defined in src/agent/index.ts:
abstract basePrompt: () => string;
abstract detectPluginUse: (response: string) => false | PluginInvocation;
abstract handlePluginOutput: (input: PluginInvocation, output: PluginOutput) => void;
pluginDetectRegex: RegExp | null = null
abstract run(prompt: string, role?: MessageRoles): void;
This allows you to flexibly define a syntax for your models to invoke plugins, handle their output and emit appropriate events, all within the context of your core agent loop.
For an example of a non-trivial core agent loop, look at the Daydreaming example. The daydreaming idea involves two agents with slightly different prompts (one biased towards idea generation and one biased towards idea filtering) and interconnects them in a converser pattern.
Daydreaming is also provided as a plugin within the framework itself in src/plugins/daydream.ts, and with smart enough prompting can be invoked automatically when the agent feels like it would benefit from some self-reflection to generate better answers.
Build-a-bot is oriented around four basic components:
- One or more models, which make API calls out to LLMs. Models never run in-process; they are always called over the network.
- Zero or more plugins, which can be used by agents to access external tools and resources.
- One or more agents, which compose models, prompts and plugins into meaningful behavior.
- One or more interfaces, which expose functionality of your agent(s) to users, such as via a websocket, Slack bot, over email or so on.
A typical implementation will expose one agent built on a range of models and prompts and able to use a range of plugins via a websocket.
Build-a-bot plugins can be either implemented specifically using Typescript (for example, the Browser, Daydream, or Shell plugins), or an OpenAPI-based plugin can be loaded from an AI Plugin JSON spec.
- Observations in the Browser plugin are frequently far too large to fit in model context, and splitting is not currently implemented.
- The Shell plugin is currently a stub.
