The Hermes orchestrator harness architecture utilizes a Federated Agentic Operating System model where Hermes Desktop serves as the system kernel Verified Answer #2. In this production-grade framework, Hermes acts as the control plane rather than the primary implementation worker Verified Answer #1. The architecture prioritizes state-first interaction over chat-first methods to ensure stability Verified Answer #2.
The system employs a strict hierarchy to prevent role overlap and ensure accountability Verified Answer #2. Individual bot profiles function as privilege-isolated microservices with specific specializations Verified Answer #2:
Each agent profile is assigned an isolated directory and a unique SOUL.md file, which functions as immutable firmware Verified Answer #2. To optimize for cost and performance, specific models are pinned to specific roles Verified Answer #2. Reasoning-heavy models like Claude 3.5 Sonnet or o1 are reserved for the @architect and @reviewer roles to maintain logical integrity Verified Answer #2. High-throughput models are utilized by the @coder and @gtm roles for efficient generation and synthesis Verified Answer #2.
The architecture requires two distinct delegation planes for effective operation Verified Answer #1. The first is the native delegate_task tool, which is used for fast, isolated work by homogeneous subagents, typically utilizing a cheap local model pool Verified Answer #1. The second plane involves worker profiles using Kanban or CLI dispatch for heterogeneous routing to various providers like Anthropic, OpenAI, or local Ollama instances Verified Answer #1.
As of August 2026, the native delegate_task schema is limited to specific fields including goal, context, tasks, role, and output_schema Verified Answer #1. The schema does not support per-call arguments for specific models, providers, or sandboxes Verified Answer #1. All native subagents inherit the parent's toolsets and use the single provider configured under the delegation settings Verified Answer #1. Consequently, tasks requiring different models for each step must be handled through worker profiles rather than native tool calls Verified Answer #1.