Autopoiesis comes from Greek αὐτo- (auto) ’self‘ and ποίησις (poiesis) ‚creation, production‘
[No tokens have been burned to write this article* 🙂 ]
In an extremely fast-paced AI/LLM world, one has to constantly vigilant about new cool patterns/tools/skills/models etc, and I sense quite a bit of FOMO (fear of missing out) by myself and some peers. AI Harness is currently the big thing, and a plethora of new harnesses or harness-extensions is coming in each day.
As a person who is very interested in how people work in organizations, I wonder what can make enable a team to be most efficient and effective. In my daily life as an engineering manager I can observe hundreds of SW engineers with diverse backgrounds and, more importantly here, extremely diverse working styles. Harmonizing anything which was not around from the beginning is an extremely tedious endeavour, and often failing or fuzzing out to some half-assed state. So how to bring reasonable AI harness to such a diverse environment? Especially thinking of corporate environments in which not everyone being AI enthusiasts.
Today I want to present a concept which was meandering through my brain for some weeks: Autopoiesis. The main thing is that it goes one step further and not only adds project/repo specific skills and context into a repo, but a fully functioning AI platform, configured for the specifics of the repo at hand. Of course, the AI platform’s source code is not part of the repo, and I don’t dare to write one (or let one write) from scratch. But with the power of containers, configs and some scripting plus semi-clever arrangements, we can add all of that with as little overhead of 20 KByte (unoptimized).
All the user needs to do: Clone the repo, prepare the config, run the docker command and then in the browser can access the AI frontend. Following example shows how I let AI add some simple hello world printing:
A cool side-effect is, that we can also „self-mutate“ the harness settings from the frontend.
This is a very first POC, and can be refined much beyond the limited first state. Many default skills and instructions could be shipped. The AI harness could be extended with MCPs connecting to a project’s bug tracker, CI, test environments and much more. Please let me know what you think about the idea? Is it worth to follow that path or is it an overhead no one needs?
* that doesnt apply to the repo. I wanted a quick POC to share the idea with a wider audience, so I used Codex with Astra for most of the way.