Recommended reading and a really cool idea.

There are a lot of interesting harness designs that are starting to emerge around tool calling and code execution. RLM is one of them. But so is this Speculative Programmatic Tool Calling approach (from the same author of RLM).

There are plenty of ways to gain efficiencies at the harness layer.

Harnesses make agents wait: the model streams a block of code, and tool calls inside it only run once generation finishes. sPTC launches the safe calls early against a copy of the environment, so tool latency overlaps with token generation instead of adding on top of it. Bad guesses get thrown away. So far it's 1 to 1.2x speedup. Very promising.