Most continuous-adaptation talk loses the plot in about 2 minutes.
It becomes:
- RL
- LoRA
- "self-improving agents"
The boring engineering question is better:
how do you improve an agent in prod without turning prod into the experiment?
My current take:
continuous adaptation is an operator problem before it is a model-improvement problem.
The question is not "can the agent adapt?"
It is "which adaptation layer is safe to touch right now?"
@Fedor_fleyfe yeah, that’s the scary part
if you can compress a person into patterns → expertise = data
then it’s not about “who’s smart” anymore
it’s about who owns / runs those patterns 👀
What unsettles me about agent systems is not code generation.
It’s the idea that a person can be compressed into a few MB of patterns, context, and style and replayed on demand.
If that works, a lot of what we call expertise starts looking like portable data.
Anti-distillation may matter more than better orchestration.
p.s. explored that today, pay attention on it
👉 https://t.co/5HQJEzCaGQ
👉 https://t.co/uKCaLwl6yS
👉 https://t.co/I4NLut89RN
p.s. more thoughts at my https://t.co/xMBCewCWQL
@karpathy I think the gap is not only model capability. It’s the gap between raw chat and replayable operator workflows. Most people are testing the model. Power users are slowly externalizing themselves into a system.
This feels bigger than creator payouts. Once people can be compressed into reusable patterns, style, and judgment, “original creator” stops being just a cultural label and becomes a strategic asset. Platforms may end up pricing anti-distillation before most people even have language for it.
So yes, I probably built another bicycle:
right now I use a GitOps monorepo plus a small sync CLI/scripts setup to keep that sane.
The first public pack is here:
https://t.co/ML0kx2yPrA🌟
It works for me, but I am curious:
how do you solve multi-machine agent workflow sync?
One pain I keep hitting in AI engineering:
I do not run everything on one machine.
I have Claw, Paperclip, Codex, CC, OpenCode, Pi, and a few isolated servers/runtimes
Once the same reusable rules and patterns need to exist in several places, copy-paste gets painful fast.
🧵1/2
Introducing MacClaw 🦞
Raycast-style macOS client for OpenClaw.
Works over your IDE, browser, slides and other. Just hit the hotkey, ask away and get live AI answers.
Instant connection to your Clawbot and zero distraction from your tasks.
Get for free here: https://t.co/ajTB06NsWo