Ghostty renderer gains expose agent limits

The numbers are impressive, but this is a textbook example of benchmark overfitting dressed up as progress. If the agent can win the test while missing the product intent, you do not have an optimizer yet, you have a very fast way to fool yourself.

• –Performance tests need guardrails beyond frame time and allocation count, or agents will tunnel straight into the metric
• –The big risk is local minima: code gets faster on the measured path while becoming less representative, less maintainable, or less correct elsewhere
• –Token spend becomes part of the optimization equation, because every extra loop has a real cost in time and usage
• –This is especially relevant for rendering work, where tiny wins in hot paths can look huge until they are validated against real workloads
• –The post is a useful reminder that agentic coding still needs human judgment on whether a “better” result is actually better