What I have been up to lately

September 30, 2026. First posted on LinkedIn on September 28; reproduced here, lightly edited, so it can be linked from places with shorter limits.

Ran git switch -c ws-project-management/coordination main
 │ devcapsule workflow claim 'Prepare fair smartgalati-test4 implementation work order'
 │ cat scripts/gcs/config.env codex-astra-site/README.md

 └ project-management: claim recorded for 12 h
  # ==============================================================================
  … +141 lines (ctrl + t to view transcript)

This is one LLM, using the tooling we developed together (me and three LLMs) to create a work order for the next LLM. Thanks to DevCapsule.

A DevCapsule project is LLM-independent: you can run LLMs in parallel on different tasks, or switch to another LLM in the middle of an implementation task, without losing any context. As a human, use (almost) any IDE you may want. You don't need JIRA, you don't need UMLs in IBM Rational Rose (anyone remember that?), as the LLMs coordinate and run project management (with your blessing, of course). They write bug reports and document design issues; from discussion to decisions, it all lives and breathes in the source tree, so every bit of context about the project is available to all the LLMs and all the humans working on it.

What's most important: the next human onboarding a DevCapsule project has to just git clone, then devcapsule project init to review and approve what the project asks to run on their computer, and then devcapsule project run. That middle step is not ceremony, it is the whole point: nothing runs on your machine that you haven't seen and said yes to. Not a base image, not a vendor download, not the host access an LLM would love to have. You don't let software you're not aware of run on your computer, and neither does DevCapsule.

What it buys you as a developer:

Onboarding in minutes, not hours or days. Clone, approve, run.
A reproducible development environment, with state-of-the-art software engineering at your fingertips. Or go your own way!
Tame the LLM. The rm -rf $HOME stories stop at the capsule's wall. Then unleash it where it belongs: writing the code, solving the problem.
Leave and come back. A month later, your IDE, your agents and your tools are exactly as you left them.

The only dependency we want on the local machine is either Linux or Windows (Mac maybe next month), a working Docker installation (use WSL2 on Windows) and a working browser. A working X desktop on Linux provides a smoother UI for the time being, but it is not necessary. I do my dogfooding mostly in the browser, with only one full-desktop instance to make sure it doesn't break.

DevCapsule is a proof of itself: it's been developed this way for the past two months, with Gemini, ChatGPT and Fable 5. More options are coming very soon (with the speed of AI development, of course).

I'll save the lessons learned for another post, because I am in the middle of fabulous learning.

Will be grateful for anyone brave enough to try it, file bugs, send PRs, of course. But more about this later.