Video: "Hermes Agent OS: Mixture of Agents + Hermes Oracle + Obsidian!" by Julian Goldie on YouTube.
What Hermes Agent OS actually is
Hermes Agent OS is the unified control layer that Julian Goldie's team built to run multiple AI agents — Hermes, Claude, and OpenClaw — from a single desktop dashboard. The appeal is operational: rather than switching between separate tools, you manage all agent activity from one place, with a shared memory system that lets each agent draw on the same context. The June 28 update extends that dashboard to include Mixture of Agents and Oracle directly, rather than requiring separate configuration for each.
The Obsidian integration, which had been a notable feature of earlier Hermes builds, sits underneath the whole system as a structured memory store. Hermes has accumulated over 3,000 individual memories in Julian Goldie's own setup — covering past tasks, user preferences, and domain context — and those memories are accessible to whichever agent is running at any point. The Agent OS update makes that memory layer visible and searchable from within the same interface rather than requiring you to open Obsidian separately.
What Mixture of Agents adds
Mixture of Agents (MoA) is Hermes's approach to the problem of any single model being wrong. Instead of routing a prompt to one model, MoA sends it to a panel — in Julian Goldie's standard configuration, Opus 4.8 and GPT-5.5 run in parallel, and Hermes aggregates their responses into one answer. The aggregation step is where the quality gain comes from: the system looks for agreement between models, uses disagreement as a signal to probe further, and produces output that reflects the best of what the panel produced rather than the full response of any single model.
The Agent OS update means you can configure and run MoA panels directly from the dashboard without editing config files manually. You choose which models join the panel, set the aggregation behaviour, and run tasks — all within the same interface you use for single-model agent work. For teams already using Hermes, this removes a meaningful friction point.
What Oracle brings to the dashboard
Oracle is Hermes's automated monitoring feature. It watches designated sources — news feeds, websites, social channels — and when it spots content matching your criteria, it drafts a response: a summary, a content piece, a social post, or whatever output you have configured. The draft arrives in Hermes ready for review rather than requiring you to notice the trigger yourself.
In the Agent OS dashboard, Oracle runs as a background process you can monitor and adjust without leaving the main interface. You can see what it has detected, what drafts it has queued, and update its source list or output templates on the fly. The practical use case Julian Goldie demonstrates is content monitoring: Oracle spots a relevant AI story, drafts a summary, and the human review step is simply approving or editing the draft rather than sourcing and writing from scratch.
How the three parts work together
The integration is the point. Running MoA, Oracle, and Obsidian memory as separate tools means managing three different interfaces and three separate contexts. In the updated Agent OS, they share a single memory layer and a single dashboard. Oracle's monitored output can feed into MoA for processing; the results get written back to Obsidian memory, which means future agent sessions have context for what was found, drafted, and approved in previous ones.
For a business using Hermes as a content and research engine, this means each cycle of monitoring, drafting and review improves the quality of the next cycle — the memory accumulates useful signal rather than discarding it. That compound improvement is harder to achieve when the tools are separate and their outputs do not connect.
What to be realistic about
Agent OS is still a developer-facing tool. Setting it up — connecting the right models, configuring Oracle's sources, building the Obsidian vault structure — requires time and some technical familiarity. The dashboard makes ongoing management much easier once it is set up; it does not make initial setup simple for non-technical users. If you are evaluating Hermes for your business, the honest advice is to treat setup as a one-off professional task and the ongoing operation as the easy part.
The memory accumulation is also only as useful as the structure you put it in. Three thousand memories is a lot, but if they are not tagged and organised in a way the agents can query reliably, the system will not use them well. That is a maintenance discipline, not a technical barrier — but it is worth knowing before you start.
Where this connects to NordSys
If tracking this kind of AI-agent news makes you wonder whether one could actually run inside your business, that's exactly what our AI Agents do — named, briefed and managed for you, no setup fee, from £6 a day.
See our AI Agents →