Unleashing Autonomy

Building An AI Agent Lab For (Almost) Free

Read all articles in the Unleashing Autonomy Series

If you are creating any form of content, be it for a newsletter, blog, email list, or for social media then you will understand the following…

“What can I write about today”

Everyone, even the most prolific writers suffer this. The trouble is, when you are stuck for an idea, you start to browse… You read X, you read Medium, you look at the newsletters you subscribe to, you watch YouTube videos… and down the rabbit hole you go.

Hours go by, and you still haven’t written a single post of value.

But what if there was a smarter solution? What if, when you turn on your computer, there is a single email in your inbox from your personal agent with a list of ideas to write about today? What about if you could pick one of those ideas, and your AI went off, wrote a draft mirroring your personal style of writing, undetectable by AI content checkers, and put it in a brand new Google doc for you to do a final edit before publishing.

All of this is possible right now, and I have been refining the process for some time.

Overnight, my agents read every new newsletter I receive, picking out interesting stories that fit my audience. They go and scan a bunch of Reddit groups I follow, browse hundreds of feeds I subscribe to via RSS, look at specific websites for new articles, browse YouTube for new videos from my favourite channels, look at Facebook pages, and monitor my social media feeds.

Anything that is interesting is summarised by my agents in a daily email, delivered before I wake up and turn on my computer. I also get a quick notification to my phone and iPad with a message that says “There are 13 things I think you should look at in your daily briefing”.

The best part about this, is you can do it all, for almost free.

You can host the agent software on your PC, a cheap Raspberry PI, and old computer you have lying around that you no longer know what to do with anymore, a free-for-life Oracle cloud server, or a cheap VPS. (I pay just over $4 a month to run the VPS powering this whole setup).

What’s more, when you have this all setup you have access to a team of workers that never sleep who can also:

• Write social media posts for you.

• Answer emails automatically.

• Act as a chat-bot on your site, or via WhatsApp, answering customers questions.

• Act as an interface to all of the worlds LLMs without paying for a subscription.

• Transcribe voice notes you record on the go.

• And automate just about any task you manually do.

… but one of the most useful tasks, which goes hand-in-hand with what I was talking about initially, is, if suddenly I get an idea for an article, let’s say I am walking the dog, listening to a Podcast, and it hits me. I fire up Telegram (how I speak to my agents, you can use other services), and ask it the following “I want to write an article on How Starlink will disrupt the cellular industry in the coming year”…

Under the hood the agents start performing their work.

• First, the master agent reads what I have written and decides how to handle my request and passes it to the appropriate agent for the task.

• In this case, the research agent is called.

• This agent then chooses the best AI models to use for my query and send off multiple agents to query various AI models.

• These models such a GTP-5, Sonar, Claude, Grok, and DeepSeek go and perform the research, searching the web, collecting the data, and returning the results. This process can take up to ten minutes, in the meantime, the Reddit agent is looking for stories on Reddit that might be relevant, the search agent searches a list of my preferred sources such as The Washington Post, The Guardian, The Financial Times, The Economist, etc. for any articles that might be relevant.

• Once all of the agents have returned their research, it is combined, and passed to the article agent. This agent will then pick a single AI to write an article based on the context of the article. If it is technical it will use Google Gemini 2,5, if it is more creative it will pick Claude Opus 4.1.

• Once the article is written, it will be passed to the personalisation agent. This is where the magic happens. This agent loads my style guide, this explains how I write based on a collection of twenty articles I have previously trained it on. It loads the writing rules, these are rules for the AI to follow so as not to be detected by AI content checkers, this is an extensive prompt with nearly forty specific rules. Finally this agent calls Claude 4.1 (this is the best tool for following all of these rules, and asks it to rewrite the article in my style. Once the article is finished it is passed to…

• The Google Docs agent. This is a simple agent that creates a new document in Google Docs and pastes the final article in.

• Once all of these tasks have completed, the master agent then send me a message on Telegram, letting me know it is done, provides a link to the Google Doc, and asks me if I want turn this into audio so I can listen to it. If I choose yes, the text-to-speak agent is called, the audio is created, and it starts playing directly in Telegram for me. When you are on the go, this is a really cool feature.

How much did it cost me for all these processes to happen? For all of the research, for the writing and re-writing… about 7 cents. To me it is priceless.

But is sounds complicated, sounds like a huge technical process which requires coding skills, but it’s not. The agent system itself will take about an hour to setup at worst case.

The agent system is all drag and drop, no coding required, and you don’t need to do any of the building yourself apart from loading the template, and entering some API keys where needed.

If this is something that you want, and you want me to write about and provide all of the templates for, then I need you to let me know, to make sure that I am not barking up the wrong tree.

Let me know in the comments.

Dan