| Runs on | The author's own machine; agent data is backed up to GitHub daily |
|---|---|
| Models | GPT 5.5 via a ChatGPT subscription; the author previously used an Anthropic model |
| Integrations | Shopify AI Toolkit, GitHub (daily backups), SKAW (the agency's Shopify insights app), HeyGen Hyperframes (video generation) |
Who built it
Jim Stewart, founder of StewArt Media, a Melbourne digital marketing agency founded in 1999 that serves founder-led e-commerce businesses. The post is a weekly update written about a month after he started using Hermes. He is the only Hermes user in the office, and it also promotes the agency's own Shopify app, SKAW.
How it works
Rather than one general assistant, the author splits the work across sub-agents with defined roles: a Shopify engineering agent, a search engineering agent with a set of Shopify-specific skills, and a client meeting prep agent. He says the agents hold context on each client's meetings, goals and roadblocks, and that each agent has its own skills that he is still building and tuning.
He describes the setup as an internal operating system for the agency, meant to improve results for clients rather than to be sold. His rule for any AI work is that it must lead to more sales for the client, and that rule shapes which tasks he gives the agents.
Changes to client sites follow a fixed sequence: diagnose the problem, propose a fix, back up the site data, then apply the fix. Backups are pushed to GitHub daily. In the example he gives, the agents traced a drop in product snippets on a client site, proposed a structured-data (schema) fix and applied it. He reports that Google's shopping crawler then showed up again in his SKAW app, and he treats this as a correlation that needs two or three more sites before it means anything.
Hermes also drafts most of the agency's social content from its own project work, without using client data, and it produced a short video from a brief prompt using HeyGen's tool. The author admits the content can still sound machine-written and the video had layout problems.
Model choice is part of the story. He runs Hermes on GPT 5.5 through a ChatGPT subscription, and notes that his Anthropic subscriptions can't be used inside Hermes while a $20 ChatGPT plan can. He wants to stay portable across models and tools, so a change in one vendor's pricing or rules doesn't break the agency's setup.
Lessons you can use
- Split work into agents with narrow roles and their own skills, instead of asking one assistant to do everything.
- Put a backup step before any change to a production site, and make the order of steps (diagnose, propose, back up, fix) explicit.
- Judge agent work by a business outcome you already track, not by how much it produces.
- Treat one good result as a lead to check on more cases, not as proof; the author himself waits for more sites before drawing conclusions.
- Keep the setup model-agnostic so you can switch providers when pricing or reliability changes.
- Expect ongoing tuning: the author calls his setup clunky and says building skills adds to his workload.
Try a smaller version
- 1
Create one Hermes profile per role (for example shopify-eng and meeting-prep) so each has its own skills and memory. See the Profiles docs.
- 2
Write the diagnose, propose, back up, fix sequence as a skill with an explicit backup step and a stop for your approval before anything is applied. See the Skills docs.
- 3
Keep the dangerous-command approval mode on for any agent that can touch a live site, so writes need your confirmation. See the Security docs.
- 4
Set up a fallback provider so the agent keeps working if your main model or subscription becomes unavailable. See the Fallback providers docs.
- 5
Start with a read-only task, such as a meeting-prep brief built from your own notes, before giving any agent write access to a client system.
Caveats
- The schema result comes from one client site; the author calls it correlation and says it hasn't been confirmed on other sites.
- The post is partly promotional for the agency's own app, and the time-saving claims (days down to minutes) are anecdotal and not measured.
- The article doesn't say exactly how the sub-agents are implemented in Hermes (for example as profiles or delegated subagents), which chat interface is used, or how client data is protected.
Original source
An original summary written from the public source on 2026-10-20. Figures and outcomes are the author’s own reports and were not independently verified.
Spotted something wrong or out of date? Tell us on X at @HermesWatcher (opens in a new tab) and we will fix it.