The Sales Machine Built on Hermes: From One Sentence to Verified Cold Email

Cold email is still run by hand: build a list, guess the addresses, hope they are right, then spend hours on it. Julian Goldie's opening line is exactly that observation, most people still stitch together five separate tools to find the right people. His answer is a machine managed from one screen, where you write a single sentence , and it finds, verifies and writes to the people you described.

The demo run is short and convincing. He types one line into the interface, SEO agencies interested in link building, and presses find and enrich. The output is real domain names, real email addresses, company size and short descriptions, all sorted into three groups : new, enriched and valid. So the demonstration is not an engine filling a table; it is the machine finding and filtering addresses on your behalf.

The machine runs inside a panel the host calls Agent OS, and beneath it sits Hermes Agent, an open-source project from Nous Research: distributed under the MIT license, not tied to a single model, and able to keep memory across sessions. The hermes-agent.nousresearch.com documentation says that memory lives in two plain text files under a strict character budget, and the skills page on the same site describes skills as separate files that follow progressive disclosure. The repository at github.com/NousResearch/hermes-agent confirms the licence and the model independence. The brain claim is therefore not marketing, it is an architecture you can inspect.

One sentence, or a list you already have

There are two doors in. If you already have a list you paste it and enrichment continues from there. If you start from zero you skip filters, spreadsheets and guessing the email format, and simply describe who you are looking for in plain English. Agencies, e-commerce stores, local services or coaches, the logic is the same and only the words change.

Finding and verification run through Hunter. According to Hunter's own help pages the tool can derive an address from a domain or find one from a name plus domain, assigns a confidence score to each address, and runs a free verification check before returning results. The verifier checks the format, filters disposable and webmail addresses, and contacts the mail server without sending anything to confirm a mailbox really exists. The catch is stated on the same pages: on accept-all domains verification cannot become conclusive, so high-confidence addresses are recommended and invented name variants are filtered out.

From filter to reply: score, write, status

After verification comes a second layer: every candidate receives a fit score out of a hundred and anything below the threshold is dropped rather than kept. Then a personal opening line and a short message are written for each name and company. The panel shows new, enriched, valid, contacted and replied as separate states, so where a campaign stands is readable at a glance. Generated text is still held behind human approval before it goes out.

What the five tabs actually cost

The old method the host compares against is familiar: one tab to find contacts, another to verify, another to write, another to send. Hours disappear and most of the time you reach people who were never a fit. Goldie says he did this work himself in the early years of his agency. seogoldieagency.com dates the company to 2018 and describes a seventy-person team running link building for B2B SaaS and ecommerce clients; juliangoldie.com spells out the workflow, prospecting link-worthy sites and then reaching out by cold email. So the comparison is not exaggeration, it is an observation from someone doing the same job.

His reason for preferring the panel over a plain terminal is a separate argument. Typing commands leaves no durable trace of the workflow, so returning later you cannot see what was done. You could write a skill file, but it becomes messy and hard to read afterwards. In the panel you can see who was contacted and when, how many messages went out, what each setting does; a campaign can be paused and irrelevant records removed. The real difference is not a prettier interface but that the system is reversible.

Shared memory and model choice

The real promise of Agent OS is not the panel but shared memory . In the host's words it is an observable vault of plain text files. When the machine learns who works for you, or when you change the writing style, that knowledge does not disappear; the content agent and research agent read the same files. The model side is equally open: complicated work goes to a strong model, repetitive work goes to a free or local one, and you switch between them with a single command. Hermes' model independence is not a competitive advantage here, it is cost control.

Who it works for

The six steps are collected here: describe, find, verify, score, write personally, send from your own mailbox. Setup is not claimed to be more complex than it is: open the app, add your API key in settings, connect an email address. The host stresses that he is not technical, yet he built it, and he answers the most common fear directly: an approval step exists and nothing is sent without you. The second warning is technical, he recommends a separate email address for sending, apart from personal mail.

The closing section is a sales call that does not weaken the thesis, it only sets the scale. The machine has just been built, the host openly describes it as experimental and says using it is genuinely enjoyable. The package, however, is not only the machine: a thirty-day roadmap, copy-and-paste prompts, four coaching calls a week and a community of more than four thousand members working on the same thing. The value here is the support layer as much as the tool.