Posted in

Nesting for your AI Agent

by stevec•

So you’ve decided to give birth to1 an autonomous AI agent, all living in your own house. Some similarities to actual new-parenting may include:

  • Nurturing something that will likely not result in you getting any positive feedback or thanks
  • Not really knowing what time it is when you’ve recreated a container for the 8th time that evening / morning.
  • Considering spending hundreds or thousands on things you didn’t previously realise were there to be spent on before the conception of the idea
  • Becoming obsessively defensive when it comes to tweaking odd pieces of open source software until 1am
  • Having to deal with output that at times, may be of different forms and consistencies
  • Being paranoid about something you’ve done which may make the whole thing die
  • Getting overly concerned about the temperatures. Always. Measuring. Temperatures. Especially if something isn’t working quite right or is making a strange noise.

So how do you prep your way to this new world? You could navigate through Reddit. You could ask your existing AI! You could ask an actual human being you trust may know what you’re on about. Or… You could read this nonsense.

How to spawn your autonomous AI agent

Whilst running an agent that does stuff2 can be set up in a multitude of way, there’s some key components needed! These depend on what you want to do but, at a minimum:

  1. A brain – this is the AI bit. The magic bit. And it is a bit magic, still. This could be run in the cloud (e.g. Claude, GPT, or DeepSeek) or you could be brave / stupid / idealistic and run the brain on your own hardware
  2. An agent harness – the software which sends stuff to the brain and runs things the brain says to run. This needs to run somewhere, like on a computer or a dedicated server / vps.
  3. An outlet – somewhere for the agent to put its output. So you know it actually did something for the money.

Let’s dive into some options for each of these…

LLM – The brain

Using cloud AI, like Claude or ChatGPT, utilises the power of goodness knows how many datacentres, chugging away in order to answer whether you should wear brogues or sandals to that gig. The underlying brain in each case consists itself of a couple of bits:

  • The LLM (Large Language Model) – for example Opus or Astra. A massive load of matrices stored in many gigabytes or terrabytes.
  • The LLM runtime software – none of the major cloud providers say what this software is, but it uses the model to calculate output based on input.
  • The User Interface (UI) – what you use to type in your questions. Not really the “brain”… this is what you see when you visit Claude or ChatGPT in your browser

For you self-hosting types, you may be running an LLM such as Gemma (from Google) or Qwen (from Alibaba) using runtime software such as LMStudio, Unsloth Studio, llama.cpp or vLLM (more on these soon!). You may or may not run a UI on top of it such as OpenWebUI.

One limitation with cloud providers is you generally cannot access the LLM brain using an API call (which is very useful for later on). All calls usually have to go through their own software or UIs.

Agent Harness

Modern iterations of the web UIs from Anthropic or OpenAI (amongst others) are agent-like; as in, they can now do things, like search the web.

Agents such as Claude Code or Codex (if you’re on a paid plan) can read and write files to your machine, make http calls, and communicate with other applications. But if you have access to an LLM via an API, you have a lot more choice. Projects such as OpenCode, Pi Coding Agent and Deepseek Harness (amongst many) do what Codex and Claude Code can do, though with the choice of where “the brain” actually runs.

But we, soon, want to look at autonomous agents3. We’ll come back to them…

An outlet: somewhere for your agent to put stuff

This is more relevant4 when we get to autonomy but when the agent actually does/creates something, we want to allow it to put that something somewhere. By default you’ll get stuff outputted in the UI you’re using (like ChatGPT or Claude, or OpenWebUI on a local setup) then you can copy and paste it somewhere else (like your CV, for example). When an agent does stuff when you’re not around, who’s there to hear the proverbial / analogous tree fall?

So the tools we allow are key here. In my agent’s case, its main outlet is a blog – so each morning I wake up to new blog posts. Another job that runs creates some content that goes into a local Gitea repo.

You could, no doubt, hook your agent up to a printer and wake up to a 90s era desktop publishing newsletter (in lieu of having a functioning fax machine at home). Perhaps you have a neat little eInk display by the front door. Or how about a message arriving on your phone, out of the blue?

All of this is possible. And all of it is in the next edition – the birth.

  1. Install ↩︎
  2. Like, stuff that’s interesting/useful/curious ↩︎
  3. Real autonomy requires some level of sentience, I reckon. So this is approximate ↩︎
  4. See footnote 3 ↩︎