The Two Big-Lab Agents You Cannot Own
Two big labs shipped personal agents this year. Both are closed. Both run on the company's computers. Both keep your memory on the company's side. And neither flagship agent is sold to Latin America.
What OpenAI announced
At DevDay on September 29, 2026, OpenAI announced Dots. Dots are persistent, always-on agents that keep working on dedicated cloud computers inside ChatGPT. They run on GPT-6 Astra. They are available to ChatGPT Pro and Business Premium tiers. The Pro tier sits at around a hundred dollars a month. Rollout is US-first.
Read that again. The agent keeps working while you sleep. It has its own machine. It is not sold to your market yet.
What Meta announced
On September 8, 2026, Meta announced Muse. Muse is a personal agent that runs on its own dedicated secure virtual machine in Meta's cloud. There is a free tier, plus paid tiers around twenty dollars a month and around a hundred dollars a month. It launched for the US and Canada.
Same pattern. Closed system. Hosted by the company. Memory kept on the company's side. Latin America is not on the list.
What both share
Four things, and they matter more than the model names.
One: closed systems. You cannot read the code. You cannot change how the agent thinks or what it remembers.
Two: hosted by the company. The agent runs on their servers, not yours.
Three: memory kept on the company's side. Your client notes, your drafts, your workflows live in someone else's database.
Four: the flagship agents are not sold to Latin America. If you run a law firm in Quito or a consultancy in Guayaquil, you are not the customer yet.
The open-source alternative you can own
There is a third option. Hermes Agent by Nous Research, MIT licensed, public repo at github.com/NousResearch/hermes-agent.
It runs on hardware you own. A laptop. A home server. A cheap VPS. That is the whole point.
It is model-agnostic. Point it at OpenAI, Anthropic, Google, DeepSeek, or OpenRouter. Or run fully local models and nothing leaves your machine.
Memory and skills are plain text files. You can open them. You can edit them. You can back them up like any other document.
It connects to 21+ messaging platforms, including Telegram, Discord, Slack, WhatsApp and iMessage. Your agent lives where your team already talks.
A scheduler runs jobs while you are away. A learning loop writes and refines its own skill files from experience. It is one of the fastest-growing open-source projects on GitHub.
Why this matters in Latin America
Most of the region still uses ChatGPT as a search box and a copy-paste tool. That is fine for drafting an email. It is not a system.
The new flagship agents are closed, rented, and US-first. You wait for access. You pay per seat in dollars. You cannot move your memory to another provider.
Owning your system means four things. Software that runs on your hardware. Models you can swap when a better one ships. Memory you can read and edit. No per-seat US pricing.
For a service business, that is the difference between renting a tool and owning an asset.
How we run it
We run this kind of agent 24/7 on our own VPS for monitoring, content and outreach. Every shipped post and every fix is a real artifact. Nothing is theoretical.
When a client site goes down at 2am, the agent notices. When a lead form breaks, the agent flags it. When a blog post needs a draft, the agent writes one and a human edits it.
That is the practical case for ownership. Not ideology. Uptime and output.
What to do this week
Pick one workflow. Lead follow-up is a good start. Write down the steps. Then decide if you want to rent that workflow from a US lab or own it on a VPS that costs less than lunch.
If you want the second option, start with the repo. Read the docs. Run it on a laptop first. Move it to a VPS when it earns its keep.
You do not need permission from Menlo Park or San Francisco. You need a machine and a few hours.
If you want a second opinion on your current site and how an agent would fit, try the free 30-second audit at softgrama.com/audit. It grades your site and shows what to fix first. No call required.