Hermes Agent setup: context limits, subagents and costs

AI Labs walks through the Hermes Agent settings its team changed after running into truncated tool output and missing details in large documents. The tutorial covers config.yaml, the Hermes config command and settings available through the desktop app. Config commands affect the active profile, so profile selection matters before editing an AI agent setup.

The speaker describes a default max bytes value of 50,000, a 2,000-character limit for individual lines and the team's decision to raise file reads to 5,000 lines. They also move the compression threshold from 50% to 0.75 to retain more context before summarization. The discussion explains the target ratio and adjustable character limits for memory.md and user.md.

For parallel work, the team raises concurrent subagents from three to five. The speaker notes the extra token use and explains spawn depth, automatic approval and assigning smaller models to subagents. Auxiliary models and reasoning effort are further cost controls; the team's example uses Gemini Flash through OpenRouter for background tasks.

Workflow coverage includes exec commands that insert terminal output into context, aliases, optional checkpointing and rollback. The tutorial also explains background notifications, a session-only system prompt and a mode that ignores user configuration for debugging. A separate sponsored segment presents Helix Canvas as a product planning platform.