Concierge is the e best way for now to involve human in the agent loop imo (github.com)

A new autonomous control system called "looop" has been introduced, aimed at enhancing agent-driven workflows by featuring a compact and easy-to-deploy architecture. This system operates as a self-contained binary, devoid of external servers or databases, and utilizes a small gated language model (LLM) to manage and execute tasks based on real-time changes in monitored environments like GitHub, Linear, or Grafana. Looop determines the most critical action to take each cycle and can autonomously spawn worker agents for various tasks. What sets looop apart is its considered approach to human involvement; it invites user input only at specific decision points to maintain a balance between automation and oversight. This innovative design addresses a common challenge in AI systems: how to effectively integrate human judgment without compromising autonomy. Looop's two-tier interaction model—steering, which allows users to set goals asynchronously, and answering, which requires user confirmation for critical actions—promises to streamline workflows while retaining essential human oversight. The system's emphasis on durability through state preservation in plain files ensures that crucial data is not lost during interruptions, making it a reliable choice for developers looking to leverage AI efficiency while mitigating risks associated with full automation.