How to use Hermes Agent to optimise C++ entropy extraction

The author ran Hermes with qwen3.6-27b locally and asked it to implement an entropy extraction technique from a friend's writeup. The instructions specified C++ for 32-bit registers, followed by a SIMD version interleaving several extractors in parallel.

The agent read the writeup, implemented the technique and wrote more extensive tests before vectorising it. During debugging, it became confused about intermediate product overflows. The algorithm uses number theory properties that make those overflows acceptable, unlike many ordinary code paths. The author instructed the agent to add comments explaining why the overflows were fine and continue; they say that mostly resolved the confusion.

The author reports throughput increasing from an initial 12 MB/s scalar implementation to about 48 MB/s. They then set a target of 100 MB/s and left the agent working. On returning the next day, they found it had stopped after six hours at just over 100 MB/s.

Reviewing the work, the author saw that the agent examined disassembly, identified its hardware and consulted online microarchitecture timing tables. It also tried approaches that did not work. The post asserts implementation correctness but does not supply benchmark methodology beyond these reported results.

Who it's for: C++ developers with a reference algorithm, executable tests and a specific optimisation target.

Prompt
Suggested prompt (written by Agents In Use, not the author's)

Source: Hacker News · nullc