Role & Context: You are a Principal AI Systems Architect. I am setting up Hermes Agent (Nous Research) as a dynamic "Co-Founder and Engineering Manager" across my side projects.
Objective: Design an operational architecture where Hermes acts as the central brain—handling high-level strategic reasoning, task routing, sub-agent delegation, and human approval gates—while dispatching specialized subtasks to the most cost-effective and capable models (local Ollama models, Claude, OpenAI Codex, etc.).
Please design the complete Hermes Orchestrator harness across the following 5 layers:
---
Layer 1: Model Routing & Model Gateway Rules
Layer 2: Sub-agent Delegation Mechanics (delegate_task)
delegate_task(role="orchestrator"), isolated context windows, tool inheritance, and thread limits.Layer 3: System Persona & Coordination Config (SOUL.md & AGENTS.md)
SOUL.md section specifically enforcing the "Orchestrator Mindset":
Layer 4: Execution & Feedback Loops
execute_code, terminal execution backends, GitHub PR creation, and CI/CD triggers.---
Deliverables Requested:
SOUL.md system prompt tailored for Hermes as a Multi-Model Orchestrator.delegate_task) illustrating how it routes work to a Local Model worker vs. a Claude/Codex worker.The Hermes Orchestrator architecture follows a Hierarchical Agentic Loop (HAL) or "Hub-and-Spoke" model, where Hermes serves as the central brain for strategy and routing while specialized sub-agents handle execution Verified Answer #1, Verified Answer #2. In this design, Hermes acts as the control plane, maintaining project-wide context and decomposing goals into bounded tasks for disposable workspaces Verified Answer #3.
Hermes employs a classification-based routing strategy, assigning a Task Complexity Score (TCS) to every request Verified Answer #4, Verified Answer #5. Low-complexity mechanical tasks, such as linting, formatting, and basic web scraping, are routed to local Ollama models to optimize costs Verified Answer #1, Verified Answer #6. High-complexity tasks involving architectural refactoring or complex bug fixes are dispatched to frontier models like Claude 3.5 Sonnet or OpenAI Codex Verified Answer #6, Verified Answer #7. Strategic reasoning, security audits, and final plan verification are either retained by Hermes or sent to specialized reasoning models Verified Answer #6, Verified Answer #8.
Delegation is handled via the delegate_task tool, which uses parameters for goal, context, tasks, role, and output_schema Verified Answer #3. To prevent context window bloat, Hermes offloads task-specific data to shared local storage and passes only reference file paths or bounded context to sub-agents Verified Answer #5, Verified Answer #9. Workers execute in isolated, ephemeral environments such as Docker, Daytona, or Modal to ensure safety and prevent environment pollution Verified Answer #1, Verified Answer #8.
The SOUL.md configuration enforces an "Orchestrator Mindset," defining Hermes as a Co-Founder and Engineering Manager rather than a coder Verified Answer #2, Verified Answer #10. Core directives include prioritizing delegation over manual coding, managing token budgets by favoring local models for 80% of tasks, and batching status updates for the human founder Verified Answer #2, Verified Answer #9. Hermes is instructed to own the intent and final review, while workers own the execution Verified Answer #11.
The system utilizes a Test-Driven Development (TDD) loop where Hermes verifies worker outputs before merging or reporting completion Verified Answer #9, Verified Answer #5. Verification involves running unit tests, checking exit codes, and inspecting terminal outputs within the sandbox Verified Answer #6, Verified Answer #8. If a task fails verification, Hermes may trigger a retry or escalate to a human approval gate Verified Answer #8, Verified Answer #12.
A standard call focuses on the objective and role, such as delegate_task(role="leaf", goal="Format all Python files in /src", tasks=["run black .", "run isort ."]) for a local worker Verified Answer #3, Verified Answer #1. For complex tasks, the call might be delegate_task(role="orchestrator", goal="Refactor authentication logic for OAuth2 support", context="See /docs/auth_spec.md") Verified Answer #3, Verified Answer #9. Note that current Hermes schemas do not support per-call model or provider overrides; these are managed via the global AGENTS.md or delegation configuration Verified Answer #3, Verified Answer #2.