How do we write code for humans with agents? Not all agents are current on the latest longevity and mortality research, but they're surprisingly good at two things: translating messy data formats and interpreting simple loss functions.
The idea here is to encode recent research into a loss function spanning seven physiological domains. The agent just has to parse your raw test data, run the script, and read a gradient table telling you which inputs have the most personal leverage on reducing the loss. The research lives in the function — the agent doesn't need to know medicine.
If you have health data (or feel like mocking up sample test results) and an agent, I'd be curious whether your agent can work with this with no hand-holding. And I'm genuinely uncertain whether the scoring decisions encoded in the function are right — some of them deviate from standard clinical guidance. Would love to hear where people think I got it wrong.
Show HN: Health optimization as agent-guiding gradient descent | Heykuki News