Agent Configuration

Build agents are AI assistants that work autonomously inside your sandbox. Configure the model, behavior, tools, and autonomy level for each agent.

Agent Settings

build.json
json
{
  "agents": {
    "default": {
      "model": "claude-sonnet-4-6",
      "autonomy": "supervised",
      "maxSteps": 50,
      "tools": ["filesystem", "shell", "browser", "search"],
      "systemPrompt": "You are a senior web developer. Write clean, well-tested code.",
      "temperature": 0.3
    }
  }
}

Autonomy Levels

Control how much freedom agents have to act without confirmation:

supervisedAgent proposes changes and waits for your approval before applyingguidedAgent applies file changes automatically, asks before running commandsautonomousAgent executes freely, reports what it did after completion

Available Tools

Enable or disable tools for agents. Each tool grants a specific capability:

filesystemRead, write, and search files in the sandboxshellExecute shell commands (npm, git, etc.)browserOpen URLs, take screenshots, interact with web pagessearchWeb search for documentation and solutionsdatabaseRun queries against attached database instances

Model Selection

Choose the AI model that powers the agent. Available models:

claude-sonnet-4-6FastHighGeneral development, iterationclaude-opus-4-6ModerateHighestComplex architecture, debuggingclaude-haiku-4-5FastestGoodQuick edits, refactoring, templates

Generation Parameters

temperaturenumber0.3Sampling temperature (0.0 - 1.0)maxStepsnumber50Maximum actions per agent turnsystemPromptstring(default)System prompt defining agent behaviorstopOnerrorbooleantrueStop execution on first error
PreviousSandboxNextEnvironment