- Author
- Posted by weirdguy, with replies from several commenters
- Published
- 8 July 2026
- Kind
- Hacker News thread, a Show HN post about an early open-source tool
- Original
- news.ycombinator.com/item?id=48833183
- Subject
- Checkable specs for coding agents
- Themes
- Saying it exactly, Handing over control
What it is
The poster, weirdguy, introduces Kastor in a first post on Hacker News. Kastor is a Go command-line tool and config language for AI agents. The post links to a GitHub repository and asks for feedback on the language design. Four top-level replies follow, plus a few nested ones.
What it says
- The poster says agent definitions end up scattered across code, prompts, tools and settings. That makes them hard to review or diff.
- The poster says the proof of concept is narrow: specs, validation, reference checks, LangGraph codegen.
- The poster says the long-term plan is Terraform-like: build, plan/apply, and state.
- Commenter empath75 says agents are not yet defined, so a spec may be obsolete soon. The poster replies that only the outer contract should be standardized.
- The poster says a Terraform provider could suit hosted agents later. It would be awkward for generating framework code.
- Commenter whattheheckheck suggests extending the idea with TLA+, and names Burr as another code target. The poster agrees to look at Burr.
What evidence it gives
The thread offers argument and a description of a proof of concept. The poster lists two runnable examples, a weather agent and a content scheduler. No benchmarks, tests or user results appear in the thread.
Interests and limits
The poster is the author of the tool and is seeking feedback. The thread is short and has few commenters. The repository and its examples were not examined for this page. Plan/apply and state tracking are described as future work, not as built features.