A runtime for a long-lived LLM agent with ambient continuous self-perception, persistent memory, defined authority, domain-specific autonomy, and forensic accountability, all in a long ongoing relationship with a human. I call this type of system an Artificial Retainer, a non-human cross between a guide dog and someone like your accountant or lawyer. It is not designed to be your friend, but it could be a valuable colleague. Think of this as an attempt to build a trusted stable agent with a stable character that could last decades.
Inspired by [1] I have been working on a system to integrate the Normative Calculus of Lawrence Becker from "A New Stoicism" [2] into an LLM agent. This works out kind of like Constitutional AI but prompt engineering based. There is a Raku and a Python implementation. I much preferred using Raku as the built in text handling and multithreading is a joy to work with. Python is very clunky compared to it. it just has better libraries.
I make no great claims for the system, it has major issues being prompt based. It is a prototype to explore the feasibility of the idea of giving a chatbot arete, a code of conduct. There are few tests, no evals so all the usual caveats! An intellectual exercise in possibilities not currently being explored anywhere else. Does it work? Hmm, almost :)
It extracts normative propositions from incoming user requests then compares then to it's own internal ethical normative propositions using the Normative Calculus. The system also uses the Decision Paradigm algorithm from Lee Roy Beach [3] to make a forecast on whether to take up the user's task or not.
I am working on an agent cognitive architecture based around LLMs. Been working on it for a while. The main difference from all the dozens out of agentic systems out there is that has been designed so that it can introspect fully and it has a machine ethics layer based on the work a modern Stoic philosopher. I am trying to build a system that I would like to exist.
https://github.com/seamus-brady/springdrift
https://arxiv.org/abs/2604.04660