I feel we are hitting the limits of existing languages that heavily rely on IDE support for context, which makes it difficult for agents reason independently. at is an experiment toward that future:

- Explicit effects: needs/using to make dependencies visible and easily mockable by agents, , enabling built-in capability sandboxing. - Zero hidden magic: No barrel files, explicit imports, and results over exceptions (encouraging minimal diffs). - Fast: Sub-millisecond cold start and ships with WASM bindings.

MCP is dead? maybe? But not tool-use, so at comes with a working MCP server to expose tools directly to agents. I’d love your feedback, particularly on the skill/tool side of things: https://github.com/aryaminus/at/blob/main/skill.md

at definitely has its limitations, and maybe it’s because the ideator is an outsider who’s not built languages before. The plan is simple, humans remain at the idea/planning side, while agents handle the shipping.

Right now, I'm focusing on building out benchmarks and finding the constraints. KISS is the mantra!

Happy to answer any questions. Leaving you with: https://simonwillison.net/2026/Jan/19/nanolang/#atom-everyth...

Additional tags: Context without LSP, Minimal diffs, Local reasoning, No macros, Capability Sandboxing