Video: "Hermes AI Agent Update is WILD!" by Julian Goldie on YouTube.
What changed in the August Hermes Agent OS update
The August 2026 Agent OS update is not a version release with a single changelog — it is a set of expansions Julian Goldie made to the full stack he has been building over the year: Hermes as the core agent, Claude as the reasoning engine, Sakana Fugu for multi-model fusion, and a shared memory layer connecting all three. The update adds new MCPs, expanded SEO workflows, voice tools, and a self-building loop that lets the stack write and test improvements to its own configuration.
That last part — the self-building loop — is the most significant addition conceptually. When the system identifies a gap in its workflow configuration, it can now draft the fix, test it against a defined quality check, and deploy it without a human writing the configuration by hand. In practice, Julian Goldie has been using this to let the Agent OS refine its own SEO prompts and content quality settings over time.
The voice tools are a practical addition rather than a headline feature. Wake word detection — say a phrase from across the room and the agent wakes up, ready to accept a spoken instruction — arrived in Hermes v0.20 Herald in August but has been expanded here. The agent now keeps a live Python session open so voice commands that require computation respond faster than earlier implementations where each command had to initialise a fresh session.
The new SEO workflows
The SEO workflow additions in this update are the most directly relevant part for a UK business owner. Hermes Oracle — the component that monitors competitor content and trending keywords — now refreshes on a tighter schedule and feeds its output directly into the content drafting queue without a manual trigger at the handoff point.
Julian Goldie also added Firecrawl as an MCP inside the stack, which gives Hermes access to live web search data inside the agent loop. Previously, Oracle used a pre-indexed data source; Firecrawl adds the ability to pull current results in real time, which matters for news-adjacent SEO topics where freshness affects ranking. To be fair, Firecrawl adds latency and cost per query — it is not something you would run on every content brief, but it is useful for time-sensitive keyword targets.
The content pipeline now includes a quality gate that checks drafts against a configurable threshold before passing them to the publishing stage. The threshold is defined in plain English — something like "this should read naturally and not repeat the same phrase in consecutive paragraphs" — and Hermes runs a separate review pass against it. If a draft fails, it is rewritten rather than published. The reliability of this depends heavily on how precisely the quality criteria are written; vague thresholds produce inconsistent results.
What is overstated and what to watch
The self-building loop is real but narrowly scoped. It works on configuration — prompts, quality thresholds, workflow routing decisions. It does not rewrite the underlying agent code. And the improvements it generates are only as good as the evaluation criteria it tests against. If you configure the quality gate loosely, the self-improvement loop tightens towards the wrong target. This is not a problem you can ignore and expect the system to fix on its own.
The "wild" framing in the title is for the thumbnail. The practical result is that running Hermes for SEO automation is significantly less manual than it was six months ago, and the August additions close the main remaining gaps in the content production side. Technical SEO still requires a separate tool or human review — the agent does not touch site structure, internal linking, or page speed.
Where this connects to NordSys
Getting Hermes Agent OS configured correctly for your specific SEO situation — your niche, your competitors, your quality bar — takes more than following the general setup documentation. The quality thresholds, Oracle targeting, and self-improvement criteria all need tuning for your context before the automation produces output worth publishing. Our SEO & AI Ranking service covers exactly this: building the right agent setup for your site rather than a generic configuration that works in demos.
See our SEO & AI Ranking service →