Local-first system to coordinate AI agents with memory, supervision and visible decisions.
The system uses Hermes as the runtime and orchestrator: agent profiles, kanban, swarms, MCP and goal decomposition. On top of that, it adds operational layers without replacing the runtime: visual workspace, Langfuse for traces, Cognee for memory, Obsidian as an editable brain and n8n for edge automation.
The architecture separates foundational knowledge, accumulated experience and future training. This means agents do not only execute isolated tasks: they can retrieve context, consult previous trajectories and improve with evidence.
Autonomy works when it has limits, traceability and human review before execution.
Operating as the studio’s internal stack. Runtime with profiles and kanban in production; knowledge and experience memory with a validated pattern; safe goal planner; observability and automation active. Some pieces, such as automatic training and differentiated model selection by role, remain documented as later evolution.
Many teams want to “use agents”, but end up with opaque automations, scattered prompts and decisions that are hard to supervise.
Hermes Agent Ops responds to a more operational problem: how to turn AI agents into useful collaborators for a studio, without losing control over what they do, when they work, what memory they consult and which decisions remain human.
The project starts from a clear position: wrap Hermes, do not replace it. Hermes already solves runtime, profiles, kanban, swarms and MCP. The studio’s work was to add what was missing to operate with judgment: memory, observability, visual workspace, edge automation and guardrails for planning goals without unleashing uncontrolled autonomous work.
This reduces unnecessary complexity and keeps each layer responsible for a clear job.
The model treats agents as collaborators that can grow. To do that, it separates three layers: foundational knowledge, accumulated experience and future training.
Knowledge lives in retrievable memory by role. Experience is documented as reusable trajectories. Automatic training remains a later phase, not a premature promise.
The architecture also comes from a real failure: a combination of autonomous dispatcher, auto-decomposition and missing limits produced thousands of tasks and token consumption overnight.
The product decision was to turn that lesson into an operating rule: autonomy starts manually, plans are reviewed before execution and any system that can originate work needs upper limits by design.
The stack combines Hermes Agent for runtime and kanban, hermes-workspace for visual control, Langfuse for traces and evaluations, Cognee for knowledge and experience memory, Obsidian as the studio’s editable base, n8n for edge automation and OpenRouter for free models with optional fallback.
The goal is not to showcase tools for their own sake, but to design an operation where every technical decision supports clarity, cost, traceability and learning.
The repo distinguishes what is in production, what has a validated pattern and what remains pending. That clarity matters: in AI systems, declaring limits is part of the product.
Hermes Agent Ops shows a way to move forward without hype: operate, measure, document failures and increase autonomy only when there is enough evidence.