overview

Agents

An agent is a conversational assistant you define alongside your ontology. It calls a language model, selected worker tools, sandboxed read and bash, and an authorized Sixb API.

You define an agent declaratively and export it from agents/; createSixb() discovers it. A worker runs it, and clients drive it over HTTP and a websocket.

Define an agent

Put each definition in agents/ and export it.

// agents/business-analyst.ts
import { defineAgent } from "@sixb/core"
import { gateway } from "ai"

export const businessAnalyst = defineAgent("business-analyst", {
  name: "Business Analyst",
  description: "Investigates customers, invoices, projects, and follow-ups.",
  model: gateway("deepseek/deepseek-v4-flash"),
  instructions: [
    "You are the business operations analyst for this project.",
    "Ground answers in the data available through Sixb, and say when data is insufficient.",
    "Prefer concise summaries with clear next actions.",
  ].join("\n"),
})

See Defining agents for every config field.

Core concepts

Concept	What it is
Definition	The agent you write with defineAgent — model, instructions, groups, selected tools, and loop limits.
Thread	One conversation with an agent, owned by a principal.
Run	One turn. Posting a user message triggers a run.
Message	A system, user, or assistant message made of text, reasoning, step-start, and tool-call parts.
Tools	Explicitly selected worker tools plus built-in read and bash in a sandbox.

Run an agent

Defining an agent needs nothing extra. Running one needs two things:

• The agent-worker process. bun sixb dev runs it for you; in production run it like the other workers.
• A sandbox factory — createSixb({ sandboxes }). The worker won't start without one, because the read and bash tools run in a sandbox.

import { createSixb } from "@sixb/core"
import { SmolvmSandboxFactory } from "@sixb/sandboxes-smolvm"

export const sixb = createSixb({
  id: "acme-corp",
  // ...storage, broker, queues
  sandboxes: new SmolvmSandboxFactory(),
})