A look at coding agent benchmarks, and what may be interesting next