What's the best programming language for coding agents?

There's a peculiar poetry in the idea that we've spent seventy years designing programming languages to be more human-friendly—more readable, more expressive, more forgiving—only to discover that a newly emergent intelligence judges them on an entirely different metric: token economy. Dan Luu's analysis (http://danluu.com/pl-tokens/) peels back the curtain on this quiet revolution. He examines how different languages are priced out in the currency of tokens when fed through large language models, and the implications ripple far beyond a simple "best language" list. This isn't about developer preference anymore; it's about the raw physics of an AI's limited working memory. Think about it as a kind of cognitive thermodynamics. Every token is a unit of thought, and the context window is the AI's universe. A language that requires 30 tokens to express an idea that another language expresses in five is effectively forcing the agent to think more slowly, to carry more baggage, to lose the thread of a large codebase before it even gets to the end. The Hacker News discussion surrounding this piece (referenced via the related link) shows the community grappling with the same sense of vertigo: the rules of the game have changed. For decades we optimized for the human eye; now we must optimize for the machine's hidden, quantized mind. What makes this genuinely weird is the inversion of values it suggests. Historically, verbose languages like Python won because they lowered the barrier for human cognition. But an AI doesn't need variable names to be self-documenting; it needs the essence of the program to fit inside its finite head. This implies a potential renaissance for terse, symbolic languages—ones that humans once mocked as "write-only." There's a delicious feedback loop here too: if AI agents become more effective in denser languages, they'll write more code in those languages, which will