Descriptions:
David Ondrej provides a full setup guide for Hermes Agent’s /goal feature — a long-running agentic loop capable of autonomously executing complex tasks for 6 to 30+ hours without human intervention. The video traces the conceptual lineage from earlier “Ralph loop” experiments through OpenAI’s /goal release inside Codex CLI on April 30, 2026, before focusing on Hermes Agent’s own implementation, which the creator argues surpasses Codex’s version on several key dimensions.
The architectural differences Ondrej highlights include a dedicated judge model — a separate AI agent that evaluates progress independently from the execution agent by reviewing the goal text and last 4 KB of agent output — plus support for mid-run sub-goals that let users steer the loop while it’s live, persistence across crashes and session restarts, and deep integration with Hermes’s existing skill library, MCP tools, checkpoints, and rollback system. The video includes a step-by-step terminal walkthrough of installing Hermes, creating a custom image generation skill powered by OpenRouter and GPT Image 2, setting credit limits on API keys, and formulating /goal prompts for autonomous slide deck generation.
Real-world use cases discussed include refactoring entire codebases across programming languages, generating PowerPoint-quality multi-slide presentations from a single prompt, and a cited example of an agent autonomously messaging potential customers to sell a farm’s excess produce. The tutorial is framed for both developers and non-technical users comfortable with a terminal.
📺 Source: David Ondrej · Published May 16, 2026
🏷️ Format: Hands On Build
