I write when I have something to say, not to keep a cadence. Every post is something I worked through — and had to explain to truly understand.
Writing since 2026
By the numbers
Agent failures often come from the system around the model. Harness engineering turns failed runs into testable changes while keeping evaluation, budgets, and permissions outside the loop.
DeepSeek reports a 47-point DeepSWE gain without changing V4-Flash’s architecture. The release points to post-training and agent infrastructure, but the public evidence cannot isolate their effects.
RecursiveMAS lets agents exchange latent states through a small trained module instead of text. In the paper's nine-benchmark setup, it used up to 75.6% fewer tokens, ran up to 2.4× faster, and improved average accuracy.
Aug
Jul
Apr