Nous Research releases Advanced Delegation plugin that lets Hermes Agent send each subagent task to a different model or profile

The plugin adds two modes to subagent delegation, and only one can be active at a time. In model routes, each named route (such as fast, deep or local) maps to a model plus either a provider or an OpenAI-compatible endpoint. In profile routes, the subagent runs as an approved Hermes profile with that profile's own model, credentials and skills. The main agent picks a route by reading one-line descriptions, and each result reports which route, provider and model handled it. [5] [8] [9]

The agent sees only route names and descriptions. It never sees model IDs, endpoints or keys, and Hermes rejects any route that isn't on the list before the subagent starts. Profile routes are the exception to this isolation: listing a profile gives subagents that profile's credentials, SOUL.md instructions and terminal access. [5] [4]

For developers, plugins register a delegation provider that either returns a route for a task or raises an error that stops only that task. The route options are added to the tool schema in a fixed order, so the prompt cache stays valid. A mode switch therefore takes effect at the next /new session, unless the user passes --now and accepts one cache break. [10] [8] [11]

Any cost savings depend on the main agent reading descriptions the operator writes. There is no trained router. Learned routers such as RouteLLM have reported more than 2× cost reduction on public benchmarks, but no published measurement yet shows how much this plugin saves. [8] [14] [4]