Hermes Agent users can cut tokens 30–50% by routing background tasks to cheaper models

Roger_M_Taylor · x · 2026-07-22

The post argues that Hermes Agent users make the same five costly mistakes and that each can be fixed quickly.
The most important mistake is using the main model for every background task; it recommends routing auxiliary jobs like compression, web extraction, and vision to a cheaper model such as google/gemini-3-flash-preview.
The author says this single change can save 30–50% of tokens.
Another issue is an empty SOUL.md, which leaves the agent on a generic default system prompt instead of a task-specific one.
The image also highlights vague goals, never checking memory, and loading all tools by default as other common failures.

More from coding & agent