1. Current "AI" is LLMs. They only work because an astronomical amount of human-written text gets their internal weights correct. A programming language "just for agents" is not going to every have phenomenal amounts of human written code. No way to train the agents in it.
2. Programming languages are for the most part, strictly descriptive and syntactically rigid right now. Sure, C has some "undefined behavior" but that sort of thing is minimal in Rust, Go, Haskell, etc etc, more modern languages. Go has almost no undefined behavior except for code that uses the package named "unsafe". Beyond that it's the very nature of programming languages to be logically undecideable and incomplete. As axiomatic systems, they have to be.
One level up, why would "AI" need a specific programming language? Can't it just use whatever language in whatever circumstance it deems appropriate? If an "AI" can figure out the vagaries of English, then it can surely write and compile an "app" in an unholy mix of SML-NJ '93, BASIC and Z80 assembly.