Why a language

Agents are programs that other programs — models — steer at run time. Nobody checks what they are allowed to do. That check belongs in the compiler.

The problem

A useful agent today is glue. The loop is Python or TypeScript, the tools are shell scripts, the browser part is JavaScript, and the device part is C. Four languages, four sets of rules, and in the middle a model whose output decides which function runs next.

Every serious agent team writes the same defences by hand: an allow-list of tools, a confirmation step before anything that spends money, a filter in front of the shell, a note in the prompt that says "never". These rules live in reviews and in prompts. A refactor that routes a tool through one more helper function quietly removes the confirmation, and nothing notices until it runs.

This is the same problem the industry had with memory and with SQL. The answer then was not more careful people; it was tools that refuse the dangerous program.

The insight

What an agent may do is a property of the program, not of the prompt. Which tools are reachable from the agent, along which paths, and whether a person approves on every one of those paths — a compiler can answer that, exactly, before anything runs. So can the question "can text from a model, or from a web request, reach a shell?" That is taint, and it is the same check whether the danger is called SQL injection or prompt injection.

A library cannot give you this. A library sees one call at a time, at run time, after the program is already deployed. A language sees every path at once, and can say no with a line number and a fix.

So in Tin, agent, tool fn, guarded, approve, capabilities and the deploy target are part of the language. The checker knows them. A guarded tool that becomes reachable without approval is a compile error. A request body that reaches SQL is a compile error. A voice agent sent to a platform without WebSockets is a compile error.

Why now

Three things changed. Models increasingly write and run code themselves, and they need a language whose errors they can act on — Tin's name the place, the path and the rule, and its safety errors say what to change. Products are turning into endpoints that agents call, so the agent is the backend. And the runtime surface finally allows one language everywhere: WebAssembly in every browser, WASI 0.2 for servers, isolates at the edge, and native binaries small enough for a device. Tin compiles to all of them from the same source.

Why us, why this shape

Tin is built by a developer who has spent years on agent frameworks and on agents that move real robots, where "the model decided to" is not an acceptable answer. The project itself is run the way Tin expects software to be made: coding agents do the work in parallel lanes, and a critic agent that never edits the code decides when a lane is done.

What we will not do

Learn Tin The memo