lbx154/Argus

What it solves

Argus addresses the limitation of current AI agents that typically stop working when a human operator goes to bed. It provides a system for persistent, long-running autonomy in research and engineering, allowing agents to plan, execute, verify, and continue work across multiple model turns without constant human supervision.

How it works

Argus uses a "Driver–Harness Model" where the AI takes the "Driver's seat" through a specialized multi-role architecture. It splits responsibilities across four distinct roles to ensure no single agent can grade its own work:

• Manager: Handles stage transitions and manages lessons learned.
• Planner: Defines the next task and the required evidence for completion.
• Engineer: Performs the actual implementation, research, and experimentation.
• Reviewer: A read-only role that provides the final verdict on correctness and evidence.

The system improves over time without retraining by using scoped Skills and Wiki findings (project $\rightarrow$ vertical $\rightarrow$ global). It integrates with various existing Agent CLIs (like GitHub Copilot, Claude Code, and Cursor) as backends to execute tasks.

Who it’s for

Argus is designed for researchers and engineers who need autonomous agents to handle complex, long-term technical campaigns that require verification and iterative refinement.

Highlights

• High Autonomy: Demonstrated a duty cycle of 95–99%, requiring human intervention only once every 310 hours over 27 campaigns.
• Role Separation: Strict boundaries between Manager, Planner, Engineer, and Reviewer to prevent self-grading bias.
• Broad Backend Support: Native integration with GitHub Copilot CLI, Claude Code, Cursor CLI, DeepSeek, and others.
• Flexible Interfaces: Accessible via a Terminal cockpit, Web UI, and mobile integrations (Telegram, Feishu/Lark).
• Customizable Autonomy: Offers different modes (cautious, pragmatic, autonomous) to control how often the agent asks for human approval.