NEW Hermes AI Goals Update is Insane (FREE)!
Hermes AI agent has released a free 'Goals' feature that enables autonomous, multi-step task execution from a single prompt. The feature uses a loop system where a judge model checks after each turn whether the goal is complete, continuing until finished or a turn budget is exhausted. Users can pause, resume, and configure the feature via CLI commands.
Summary
The video covers a new free update to the Hermes AI agent called 'Goals' (or persistent goals), which allows the agent to autonomously pursue a defined objective without requiring repeated user prompting. The feature was inspired by Ralph Loops, a concept popularized with Claude Code, and was announced by one of Hermes' co-founders.
The core mechanic works as a loop: after each turn, a separate 'judge' AI model evaluates whether the goal has been completed. If not, Hermes continues working automatically. This loop persists until the goal is achieved, the user pauses or clears it, or a configurable turn budget (defaulting to 20 turns) is exhausted. The judge model and max turns can both be customized in the Hermes config file.
Users activate the feature via terminal commands: `/goal [text]` to set a goal, `/goal status` to check progress, `/goal pause` to pause, `/goal resume` to continue later, and `/goal clear` to end the session. A key feature highlighted is persistence — users can close their terminal, and the next day run `/goal resume` to pick up exactly where the agent left off.
The presenter demonstrates the feature live by instructing Hermes to build and deploy an SEO-optimized blog about Hermes AI, deploy it to Netlify, and continue posting daily until it ranks. The agent is shown generating HTML files, writing blog posts, and deploying the site autonomously.
The presenter recommends using specific, clear, and measurable goal prompts, and suggests use cases including content production, research reports, code debugging, feature porting, and CLI tool building. A deeper guide and 30-day roadmap are promoted via the AI Profit Boardroom community.
Key Insights
- The Goals feature was inspired by 'Ralph Loops' — a looping agent architecture popularized with Claude Code — where a supervisor/judge model determines after each turn whether the task is complete, and if not, the agent continues automatically.
- Hermes Goals includes a turn-based budget (default 20 turns) to prevent the agent from running indefinitely out of control, and this limit can be changed via the 'goal.max_turns' setting in the config file.
- The presenter demonstrates that goals are persistent across sessions — a user can close their terminal, come back the next day, and run '/goal resume' to continue the agent exactly where it left off.
- The judge model that evaluates goal completion is configurable — users can change both the provider and the model used for the evaluation step inside the Hermes config, giving control over how goal success is determined.
- The presenter shows a live demo where a single goal command — instructing Hermes to build an SEO-optimized blog and deploy it to Netlify — results in the agent autonomously generating HTML files, writing blog posts, and deploying the site without further user input.
Topics
Full transcript available for MurmurCast members
Sign Up to Access