Hermes Agent: subagents vs teams and Kanban setup

This tutorial compares ways to coordinate work in Hermes Agent: temporary subagents, goal loops, persistent bot profiles and Kanban task graphs. The speaker enables task delegation, steers a running helper and compares research responses with and without delegation. The reported parent context usage falls from 102,000 to about 37,000 tokens, but excludes the helpers' token usage. It is not a total usage saving.

The goal mode demonstration sets completion criteria, then adds a requirement for peer-reviewed sources while work continues. The speaker shows a default limit of 20 turns and warns that extended loops can consume model quotas. Model routing options include assigning a local model to bounded work while reserving synthesis and judgment for a stronger model. Local-model routing is discussed as an option; the shown research uses hosted models.

Bot profiles retain their own skills, memory and working context. The Kanban setup adds dependencies, review stages and handoff records stored in SQLite. Viewers learn how the dispatcher, triage specifier and task decomposer move cards through the board, and how individual workers can use goal loops.

In the speaker's small research test, delegation and orchestrator-led decomposition produced similar quality and token usage. Automatic decomposition and goal loops added overhead. The speaker explicitly says this is not a proper benchmark; the choice depends on the workflow's need for continuity and inspection.