Module eight of Hermes’ Agent Masterclass introduces sub-agents and delegation to overcome single-agent limitations by enabling parallel task execution through isolated child agents that receive specific goals and contexts, improving efficiency and reducing costs. The module demonstrates managing delegation depth, execution modes, and tool restrictions, culminating in a multi-level delegation example that highlights Hermes’ flexible and powerful autonomous workflow capabilities.
In module eight of Hermes’ Agent Masterclass, the focus is on sub-agents and delegation, which addresses two main limitations of single-agent loops: context window overload and serial processing slowness. Delegation allows a parent agent to spawn child agents with isolated contexts that work in parallel, returning only concise summaries to the parent. This approach keeps the parent’s context clean and speeds up task completion by distributing work across multiple agents. The key tool introduced is the delegate task tool, which assigns specific goals, contexts, and scoped toolsets to sub-agents, enabling efficient and isolated task execution.
A sub-agent is essentially a fresh agent instance that receives a brief from the parent agent and works independently with its own conversation, tools, and iteration limits. The parent only receives the final summary, preventing intermediate data and errors from cluttering its context. This isolation is crucial because sub-agents have no memory of the parent’s history and only know what is explicitly provided in their brief. Properly defining the goal and context is essential to avoid delegation mistakes, and limiting the sub-agent’s toolset helps maintain control and security by preventing unauthorized actions like recursion or message sending.
The video also highlights the cost-saving benefits of delegation by running sub-agents on cheaper models while reserving stronger, more expensive models for the parent agent’s orchestration and synthesis tasks. This setup was demonstrated through a research task comparing a single-agent approach versus a delegated multi-agent approach, showing that delegation can reduce costs by about five times without sacrificing quality or significantly increasing completion time. Parallelization is managed through batch concurrency, with a default of three concurrent sub-agents, and results are returned in a stable order.
Control over delegation includes managing the depth of delegation trees, with a default maximum spawn depth of one to prevent runaway recursive delegation. However, deeper delegation trees can be enabled by assigning the orchestrator role to child agents and increasing the spawn depth, allowing multi-level task breakdowns. Execution modes include synchronous delegation, which blocks the parent until completion, and asynchronous delegation, where sub-agents run in the background and return results later. The choice of delegation versus other tools like cron jobs or single loops depends on the nature of the task, emphasizing the importance of selecting the right approach for efficiency and durability.
Finally, the module concludes with a demonstration of a three-level delegation tree for a complex research task on machine learning in sports, involving multiple orchestrators and sub-agents working in parallel and synthesizing results at each level. This example showcased the flexibility and power of Hermes’ Agent delegation system, including configuration adjustments for spawn depth and iteration limits. The module wraps up by previewing the next topic on profiles and durable task collaboration using the Kanban board, which will further expand the capabilities for advanced autonomous workflows in Hermes’ Agent.