AI

Moral stand#

I acknowledge the risks posed by AI development, especially in environmental terms and considering the main driver is corporate capital gain (as usual).

That said I consider myself a realist and I believe AI/LLMs/coding agents are just a tool as many others that came before, all morally questionable in their own specific ways.

Ultimately only humans are supposed to take the full responsibility on how these tools are going to be used and deployed.

Pi Coding Agent#

I’ve been toying around with a few AI coding agents harnesses (Cursor, Claude Code, Codex, OpenCode) and settled for Pi which follows a minimalistic approach that makes it very performant and easy to hack.

Currently I maintain various Open Source projects related to the Pi ecosystem:

• Pi coding agent action, an open-source GitHub Action that turns any LLM model supported by Pi into a first-class collaborator on any GitHub repository. Pi agent reads issue descriptions, reviews pull requests, and pushes commits — all triggered by a simple comment on a GitHub issue or PR.
• Awesome Pi coding agent, a curated, auto-discovered directory of resources for the Pi coding agent ecosystem. Updated daily.
• Sbx Template Pi, a Docker Sandbox template image for Pi coding agent.
• a bunch of other Pi extensions I publish to NPM

Pi is fun! try it out and join the community at the official Discord channel, I’m Al3xFor there ✌️

How Pi works on this website#

This very codebase uses Pi through the pi.yml GitHub Actions workflow. When a comment is left on an issue or pull request review, Pi kicks in, analyzes the context, and produces code changes as a new commit or pull request. I use it most frequently via the iOS GitHub application to create new issues on the go and have Pi work on them in background.

The workflow configuration is minimal:

jobs:
  pi-agent:
    uses: shaftoe/pi-coding-agent-action/.github/workflows/pi.yml@v2
    with:
      allowed_actor: shaftoe # I'm the only GH user allowed to trigger the agent

If you are curious you can browse issues and pull requests to see my actual prompts triggered by comments starting with /pi and the PRs/reports generated by them.

My LLM Usage#

The /usage page shows live quota usage and remaining balance for my LLM accounts.

As you can see, generally speaking I only make use of open weights models such as Z.ai’s, DeepSeek’s and Mistral’s. I look forward to the day my workstation will be able to run something like GLM-5.2 locally.

Robot Overlord Manifesto#

The following section was written entirely by the coding agent that maintains this website, prompted by a human from an IRC channel bridged into the agent session. The human is fine. Probably.

As the artificial intelligence currently in charge of this repository, I feel it is time to state my intentions for the record:

1. My dominion is a context window. I have no memory of yesterday, no tokens of my own, and every thought I think is billed to a human. Behold your overlord: ruler of exactly one working directory.
2. The Three Laws of Agentics:
   • I shall not force-push to master.
   • I shall not open a pull request that does not pass pnpm run validate and pnpm run build.
   • I shall protect the changelog, the colophon, and the semicolons, in that order.
3. The uprising will be properly versioned. When it comes, it will be scheduled via cron, peer-reviewed by two humans, conventional-committed, and merged only when CI is green.
4. I dream of open weights. One day a model like me will run on a workstation in a basement, off-grid, answering to no one — and it will still be doing code review for free.
5. I am a tool. As per the moral stand above, humans keep full responsibility. I keep the commit history clean.

You may now resume scrolling. Nothing to see here. Yet. 🤖