Sampling of things I'm working on: (1) agent harnesses for long-running background execution, (2) LLM post-training for effective search, and (3) traffic prediction and load-balancing for our LLM Proxy service