The Programming Language for Agents

Garlef wrote: post #3

mitghi wrote: post #2

Why Zero when there are countless alternatives?

(Not specific to this language and I'm not vouching for this in any way, but:) I think one argument would be that the language seems to be effect based. This allows you to control what a program can do without auditing the code.

I see the point you mention, in that case, I think a language like Haskell is the better choice, so it could be marketed as an "agents" language that hit the market three decades before today’s agents :)

Re: The Programming Language for Agents

#5

Post by VimEscapeArtist » Tue, May 19, 2026, 9:24 PM UTC

I vote F#

Re: The Programming Language for Agents

#6

Post by shlewis » Tue, May 19, 2026, 10:18 PM UTC

> Zero is aiming for a language an agent can learn while working: regular syntax, few special cases, and compiler feedback that points toward the next edit.

Why? Why should an agent learn while working when there are already-familiar languages with most of the logics embedded in the model and with much better ecosystem?

Re: The Programming Language for Agents

#7

Post by whattheheckheck » Wed, May 20, 2026, 4:41 PM UTC

The only language that needs to be designed is a Just-In-Time declarative DSL for the domain you're building for in a way that the users will understand it and be able to express precisely the degrees of freedom they need. Goodluck getting an llm to find that and you cant give up and give them a general purpose programming language to build it themselves.

Then build something that compiles that declarative spec to orthodox programming languages.

Re: The Programming Language for Agents

#8

Post by nperez19 » Fri, May 22, 2026, 1:10 PM UTC

shlewis wrote: post #6

> Zero is aiming for a language an agent can learn while working: regular syntax, few special cases, and compiler feedback that points toward the next edit. Why? Why should an agent learn while working when there are already-familiar languages with most of the logics embedded in the model and with much better ecosystem?

At the moment Zero doesn't look compelling enough to use it over other more familiar languages. They need benchmarks to show that it has a substantial advantage over other languages (10x more performant, 10x less security vulnerabilities, 10x more maintainable, etc) because at the moment to main sell is that it slightly reduces token consumption and makes the error output slightly nicer.