Notes /
Shipping an agentic kit as Markdown
Packaging an agent workflow as a plain repo for agents to read, and what that suggests.
I really enjoyed OpenAI’s article on harness engineering, and I built an agent workflow I’m pretty happy with based on Codex CLI and Linear.
Which led me to the next question: how do I package and reuse this across my other projects?
Copy-pasting SKILLS everywhere is tedious and kills maintainability. The standard npm/CLI route was overkill, as the main consumer of this repo is an agent, and the whole thing is just Markdown docs - way too much hassle.
So I tried something simpler - moved it into a plain GitHub repo with the README written specifically for agents to consume. Added lightweight version control on top and it just worked.
But that isn’t the real point. The whole thing led me to the thought that in this new reality, instead of code-based deterministic packages, we could start shipping probabilistic descriptions that get constructed on the fly, shaped by the agent’s and user’s constraints and considerations in real time - maybe this is Software 3.0 Karpathy was talking about.
You can check out the repo here if you’re curious:
Also posted on LinkedIn ↗