Builder Urges Composable AI Systems For Future-Proof Development
A bit vague for sure, but just think of it as building for the new and for the future. You should just be able to leverage primitives, swap them out, and iterate on them.
Two in particular that many devs are not thinking enough about in terms of customizability are LLMs themselves and Agent Harness. You don't want to be locking yourself up with one specific model or harness. You need to own that layer where things are headed. And some might complain that it's too much work. Actually, it's getting easier to tune LLM directly with an agent (e.g., autoresearch) and develop your harness with things like Pi.
This is how you get those compounding effects. If there is enough interesting, I might convert this entire thing into a more comprehensive article with more insights, data, and research.
elvis@omarsar0For future-proof, build AI that's composable. Regardless of what you use, all these should be composable, iterative, and customizable: - LLMs - Evals - Automations - MCP/CLI tools - Skills/Memory/Context - Agent Harness (Codex, CC, Pi,...) The compounding effects are insane.