The main session of the Hermes Agent is clearly equipped with a fallback, but the sub-agent still directly reports the provider error, which is not an accident. The official Fallback Providers documentation has been marked separately: Subagent delegation does not inherit fallback configurations. In other words, the master agent will be automatically downgraded, which does not mean that the child agents generated by delegate_task will also use the same set of fallbacks.
What do you do?
- If the subtask is important, specify the provider/model for the delegation separately, and don't rely only on the main session fallback.
- When cost sensitive, pin sub-agents to a cheaper, more stable model.
- The high-risk task confirms that the provider is healthy in the main session before sending it to the sub-agent.
Why Hermes designed it this way
Because sub-agents are more like independent execution units. The official documentation allows you to route the delegation separately, but you don't want the subtasks to silently follow the main session's fallback logic, resulting in uncontrollable cost, speed, and behavior.
So don't stare at "fallback is obviously open" on this question. You need to think differently: the fallback of the master agent is the master agent, and the sub-agent either specifies the provider separately or accepts that it is not automatically downgraded. Think about this boundary, and many problems that look like random errors will be solved.
Official open source address: https://github.com/NousResearch/hermes-agent; Official document entrance: https://hermes-agent.nousresearch.com/.
