My approach is https://tix.im/ - not launched yet. But from expedience: modeling everything as workflows of composeable nix flakes (expert systems from programs, skills, agents, workflows and their snippets, DBs blobs, service descriptions, ...) and then run everything in highly isolated environments. It is not only so much more efficient, reduces errors, increases quality, reduces unnecessary cognitive load, ...
I start to move more and more daily tasks into the systems. Even UI programs become interactive workflow steps, which allow me to attach from anywhere back to a task in an environment optimized for the task.
Dependency updates for example, a workflow. A subgraph from a security scanner expert checks each change in the dependency tree.
tix.im combines conecepts of expert systems with LLMs and other AI systems. Everything is composable via nix flakes. Any change is modeled via workflows that run in isolation. Each step type, is composed from the nix flakes defined and installed in the environment. Not only the projects nix flake (devshell), but also the tix project enabled experts. vx-maintainer is my software development expert - does only know about conecept of software development. Each workflow implemented is steps and workflows. For example, each change lands via land-feature workflow, it has ~40 steps. Workflows have complex parameters that define the flow inside the workflow DAG. In this case, rebase is mechanical, if it fails, medium smart rebase fixer is spawned, then mechanical step checks if all commits arrive in the rebase. Then it builds a collections, spawn a review swarm, open-code-review, etc... everything is tracked in an own issue tracking system and workflows can be restarted with minimal overhead.
Then I have an optimized queue-manager, an own agent prompt with an goal that can never be fulfilled. He manages that items from the tracker get done and land, extra workflows mechanically later check the landing.
rust expert, knows about rust specifics, has snippets, agents, skills etc. security expert has workflows with red-teaming, cyber models that get registered at the internal LLM gateway for usage, ...
I start to model everything in my daily live as an expert system now, with workflows that can be used by other expert flakes.
Workflow steps can be shared and also run wayland and x11 apps, rmux multiuser sessions etc. So, my goal is to move most of my tasks into some specialized environment.
And this is just a small part of the architecture - it is fully decentralized and user owned.
>But from expedience: modeling everything as workflows
No thanks. I like private computing. I like all the fun stuff you can do with Computers. I'll never want agent based computers.