Skip to main content
The goal agent runs multi-step tasks autonomously — set up environments, debug, refactor, and run tests with a step budget and security gates.

Usage

From zsh/bash (no fish required):

Engines

How it works

The goal agent uses a Butterfish-style run/fix loop with shell history, file reads, and security gates. Each step can invoke any registered skill or safe shell command. Use -y to auto-confirm risky steps only when you trust the environment.
  • agent_plan — upfront JSON planning (separate from reactive goal loop)
  • arka loop — alias for arka goal when GOAL_ENGINE=auto