Hermes /goal is insane… just watch

Key Concepts

• /go (Slash-Go): An autonomous agent feature that executes complex, multi-step tasks until a defined end-state is reached.
• Hermes Agent: An open-source AI agent framework featuring a built-in "Judge" model, persistence, and sub-goal management.
• Judge Model: A specialized, unbiased AI agent that evaluates progress, ensures the execution agent stays on track, and determines when a task is complete.
• Codex CLI: A command-line interface tool often used as a runtime or "CTO" agent for executing code, file edits, and sandbox operations.
• Autonomous Loops: The evolution from simple "Ralph" loops (infinite, unguided) to goal-oriented, judge-monitored execution.
• MCP (Model Context Protocol): A framework allowing agents to interact with external tools, skills, and data sources.

---

1. The /go Feature: Mechanics and Advantages

The /go feature represents a shift from simple, repetitive AI loops to goal-oriented autonomous execution. Unlike earlier "Ralph" loops—which lacked a stopping condition and often wasted tokens—/go is designed for efficiency and precision.

• The Judge System: Hermes uses a separate, unbiased AI agent to review the last 4KB of execution logs against the system prompt and the user's goal. It decides whether to continue, pivot, or terminate the task.
• Persistence: Tasks can survive crashes and restarts. Users can resume a goal exactly where it left off.
• Mid-run Interaction: Users can inject "sub-goals" or steer the agent while it is running, allowing for real-time course correction during long-duration tasks (6–24+ hours).
• Built-in Integration: Unlike manual setups, /go is native to the Hermes harness, inheriting features like checkpoints, V2 rollbacks, and LSP (Language Server Protocol) listing.

2. Step-by-Step Setup Process

1. Infrastructure: Deploy a VPS (e.g., via Hostinger). Use SSH to access the server terminal.
2. Installation: Use the Hermes one-liner installation command from the official GitHub repository to install dependencies.
3. Authentication: Configure Hermes to use a ChatGPT Plus/Pro subscription via the provided authentication URL and 9-character code, allowing for free usage of high-end models like GPT-5.5.
4. Configuration: Set up API keys (e.g., OpenRouter) using Hermes config set [key_name] [value] to ensure keys are stored securely in the root directory without being exposed in the context window.
5. Skill Creation: Define custom skills (e.g., image generation or slide deck creation) by providing specific prompts that the agent saves to its library for future use.

3. Real-World Applications and Case Studies

• Business Automation: A user with excess farm produce used the /go feature to autonomously message potential customers and facilitate sales, directly increasing revenue.
• End-to-End Testing: Developers use /go to autonomously test every feature of an iOS app in an emulator, running for hours to ensure full coverage.
• Data Backfilling: A user automated the backfilling of 17,000 records while sleeping, saving hundreds of hours of manual labor.
• Content Creation: Hermes can be tasked to generate professional, editable PowerPoint presentations by fetching reference styles from Midjourney and using GPT-4/5 models to structure the content.

4. Framework: Hermes as CEO, Codex as CTO

The most powerful configuration involves a hierarchical agent structure:

• The CEO (Hermes): Orchestrates the high-level strategy, manages the /go loop, and delegates specific technical tasks.
• The CTO (Codex CLI): Acts as the sub-agent responsible for the "dirty work"—writing code, managing Git repositories, and deploying applications.
• Methodology: The user provides a clear, measurable outcome (e.g., "Build a NextJS reporting app with specific deliverables"). Hermes spawns Codex sub-agents to execute the development, marketing, and deployment phases autonomously.

5. Key Arguments and Best Practices

• Definition of Done: The success of /go depends entirely on the clarity of the "end objective." Vague goals (e.g., "build a great app") lead to inefficient loops; measurable goals (e.g., "tests passing," "10 customers acquired") allow the Judge model to function effectively.
• Security: Always use environment variables or secure config commands for API keys to prevent them from being sent to the LLM provider during execution.
• Efficiency: Use "Fast Mode" for priority processing, though be aware this consumes credits at a higher rate.

6. Notable Quotes

• "The best answer often comes from the harness, not the weights." — Highlighting that the agent's framework (tools, judge, memory) is more critical than the raw intelligence of the underlying model.
• "If you don't have real actionable use cases, what's the point?" — Emphasizing the need for practical application over theoretical experimentation.

Synthesis/Conclusion

The /go feature is a transformative development in AI agent technology, moving the industry toward "outcome-based" computing. By combining a robust "Judge" model with persistent, multi-agent orchestration (Hermes as CEO, Codex as CTO), users can automate complex, multi-hour workflows that were previously impossible. The key to success lies in defining clear, measurable deliverables and leveraging the agent's ability to learn and store new skills autonomously.