3 Comments
User's avatar
Pawel Jozefiak's avatar

The latency compounding problem is real. I've been running an autonomous agent (Wiz) that builds apps during nightshifts, and when I didn't optimize the loop steps, a 2-second model call became 30 seconds of waiting across 15 iterations.

The sandboxing insight resonates. Cursor treats execution infrastructure with the same rigor as model development - that's what separates production agents from demos. When I analyzed who's actually winning in the agent space (https://thoughts.jock.pl/p/ai-agent-landscape-feb-2026-data), this infrastructure-first approach was the common thread.

Curious: how are you handling tool usage - baked into training or prompt-based?

Milo's avatar

Great deep dive on the systems engineering behind agents. The 'diff problem' section is particularly insightful - it explains why so many early AI coding tools felt unreliable even when the underlying model was decent.

The MoE + speculative decoding combo for latency is clever. Code's structural predictability makes it ideal for aggressive speculation. Curious if this approach will become standard as more teams train coding-specific models vs. fine-tuning general purpose ones.

Diego Bonifacino's avatar

Creatism embodies the spirit of shipping. It's not just about ideating, but executing and getting work into the world. Shipping is the ultimate creative act.