ukanwat/selfstarter

What it solves

SelfStarter is designed to move AI agents from a "chat-based" management style to a "brief-based" style. Instead of a human providing step-by-step instructions (like managing a junior employee), the agent is given a high-level goal, constraints, and a definition of "done" (like managing a senior employee), and then works autonomously over multiple sessions to achieve that goal.

How it works

The project provides a harness that allows a coding agent to interact with a real-world software environment—specifically Unreal Engine 5—via the Model Context Protocol (MCP). The agent uses a set of tools (including viewport capture for "eyes" and Python scripts for engine control) to plan its work, execute changes, and self-verify its results. The harness manages the lifecycle of the agent and the editor, automatically relaunching the editor if it crashes and resuming agent sessions to maintain continuity of the plan across cold starts.

Who it’s for

It is intended for developers and researchers interested in autonomous AI agents, long-horizon execution, and the capabilities of LLMs to perform complex engineering tasks in professional software tools without human intervention.

Highlights

• Autonomous World Building: Capable of building a playable city in Unreal Engine 5, including geography, traffic, and game screens, without human guidance.
• Agent-Agnostic Harness: Supports various agent CLIs (e.g., Claude, Gemini) as long as they have shell/file access and can reach an MCP endpoint.
• Self-Verification: The agent can use viewport captures and screenshots to notice and fix its own mistakes.
• Long-Horizon Execution: Designed for work that spans multiple sessions where the agent must maintain its own continuity through documentation.
• Robust Supervision: Includes scripts to monitor agent health and manage editor crashes with exponential backoff.