Sergio rides shotgun

At 9:30 AM on the first of last month, my phone buzzed with a Telegram message from a contact I had created six weeks earlier. "Monthly tax return is due. Want me to handle it?" I sent back "do it." A few minutes later, the reply came: a transaction number from the federal tax authority and a status of "filed." I hadn't opened a browser, typed in any ID, or been at a desk. That's Sergio doing his job.

Sergio is what I named my deployment of Hermes Agent, Nous Research's open-source agent. It runs on a Fly.io app called hermes-fede, with a single dedicated volume, and I talk to it through a Telegram bot. Voice memos get transcribed. Files come back as files. There is no laptop in this loop. The whole point of putting an agent on a remote server with a Telegram gateway is that it doesn't care where I am or what I'm doing.

How the tax skill got built

The first time I had to file my monthly Argentine cross-province gross-income return through SIFERE (the federal system for taxpayers operating in more than one province), I drove the browser myself and narrated each step to Sergio as I went. At the end I asked him to save what we had just done as a skill. He wrote a markdown file under skills/argentina-tax/sifere-monthly-filing:

---
name: sifere-monthly-filing
description: File a monthly cross-province gross-income
  return via Argentina's SIFERE web portal.
---

## Step 1. Sign in
1. Open the federal tax authority portal.
2. Sign in with the taxpayer ID and password.
3. From the services menu, click the SIFERE Web entry.

## Step 2. Open the monthly filing list
1. Click "Monthly returns".
2. Identify the period that still needs to be filed.
3. Click "New monthly return" and pick the year and month.

## Step 3. Fill in jurisdiction values
1. For each jurisdiction, calculate the total and the deduction.
2. Confirm the totals before submitting.

The next month, Sergio filed it himself

The next month, the cron-triggered reminder fired in the morning. I sent back a single message: "do it." Sergio loaded the skill, opened a real browser session, signed into the tax portal, opened the new period's filing, copied the previous month's structure, signed it, and submitted. The reply was a transaction number and a status of "filed." The agent doesn't get smarter; the skill file does, and the agent reads it. That's the closed loop. Every successful run becomes a paragraph in a markdown file the next run consults.

The labor package at the toll booth

Different story. I was on Highway 5 heading to Córdoba when the call came in. The woman who looked after my grandmother had quit that morning, four years on the job, no notice. I held the wheel with one hand and the phone with the other and recorded a 1:42 voice memo to Sergio:

"Andrea resigned today. Four years on the books as a registered employee. I want a resignation letter, a settlement document, and a guide on what I owe her. Do it in Spanish, of course. I'm driving, can't type."

Sent. Eyes back on the road.

By the time I rolled up to the next toll booth, three files were sitting in the chat, drafted in Spanish: a resignation letter for Andrea to sign, a settlement record, and a step-by-step guide on what I owed her. The guide cited Article 232 of Argentina's Labor Contract Law (Law No. 20.744), the article that governs compensation in lieu of notice when an employment relationship ends abruptly. Sergio had asked me five clarifying questions before drafting (was she registered formally with the tax authority, written or verbal contract, how many years, what I was worried about legally, anything still owed), gotten my answers, then web-searched current resignation templates and wrote the package against the actual statute. The cost wasn't the documents. It was that I didn't have to open a laptop or call anyone for two hours. I read them with the engine idling and the boom gate ahead. Good enough that my next call was to my accountant to confirm, not to redo.

Explaining LLMs to grandma

A few days later my grandmother asked me what one of those models everyone is talking about actually is. I asked Sergio for a Spanish explanation she could follow, then to read it aloud. The text used the "someone who read millions of recipes" analogy: not understanding cooking but recognizing why vinegar reacts with baking soda. He generated the audio with OpenAI's text-to-speech instead of Edge, because the Spanish prosody is noticeably more natural, saved it as an MP3, and that's what she listened to. She said it sounded like a person.

What I sell

Around the same time, friends started asking me to set theirs up. What I sell is the whole package: the deployment on a remote server, the connections to the services they actually use (Gmail, calendar, web search, voice synthesis), the routing and metering of the language-model tokens on their behalf, and a Telegram bot already wired up to all of it. The customer never opens a config file. By the time I hand the bot over, everything is set up. They just send messages. USD 50 a month for the base tier, 90 with the email connectors and ElevenLabs voice. The Spanish landing page at fedesapuppo.com/hermes pitches it at accountants, lawyers, and small-business owners, because those are the people I had stories for. The product isn't the model and it isn't the agent. It's the wired-up deploy.

What changed in how I work

• Telegram is a superior interface. Your phone is almost always in your pocket; your laptop isn't.
• The unit of value is the skill file, not the conversation. After every meaningful task, ask the agent to write down what it learned.
• A cloud-resident agent on a VPS beats a faster local one. Persistence is the feature.
• Honesty about failure is what makes an agent trustable. A magic-feeling agent that lies about what it did is worse than a slow one that reports back.

The drive from Alta Gracia to Córdoba is still 36 minutes. The difference is what you can ship in those 36 minutes if you've taught your phone to be useful.

---