03 Oct 2026 · 4 min read
agents
SIFT: What It Actually Costs to Make a Coding Agent Improve Itself
MIT and Sakana AI's SIFT framework beat the prior self-improving-agent SOTA by using a Bradley-Terry judge instead of more patches. I ran the real cost numbers — $150 and under 5 hours wall-clock for a 4.4-point jump on Polyglot.
Read more





