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

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(),
})

For each turn, the worker offers that agent's selected tools plus read and bash, executes them in their respective boundaries, and persists the reply. See Running and streaming.

Related

• Defining agents — the defineAgent config.
• Running and streaming — threads, the HTTP API, and the websocket.
• Tools and gateway — selected worker tools, connector-backed web access, sandboxed file and command access, and scoped data access.
• Authorization — gate who can use an agent and what it can reach.
• Sandboxes — where the sandbox tools run.