Tin: A GC-Free Systems Language Written Almost Entirely by AI Agents

Tin is a compiled language for servers and tooling that went open source a week ago at v0.3 and has already merged 354 PRs from nine contributors. Its distinguishing feature is process, not syntax: the compiler, runtime, standard library and docs are written almost entirely by AI coding agents working from design documents, with humans deciding scope, reviewing and merging.

Under the hood, Tin self-hosts its own compiler (~46k lines), verified by triple-bootstrapping to byte-identical binaries across macOS, Linux arm64 and Linux x86-64 in CI. It ships its own assembler and linker (no clang, ld, or libc), producing static PIE Linux binaries. Instead of garbage collection, it enforces a region-based memory model at compile time, combined with a thread-per-core, share-nothing concurrency model. Server semantics such as deadlines, memory and task limits are first-class language constructs, errors are modeled as 'faults' that cannot be silently ignored, and query strings are typed to prevent SQL injection at compile time. The project also includes a self-written HTTP/2 and gRPC server, plus a from-scratch TLS 1.3 stack with post-quantum hybrid support.

Benchmarks against Go show throughput gains in cached and mixed workloads but a memory overhead and a database p99 latency disadvantage. The crypto stack is unaudited, there are known x86-64 backend and Kafka client bugs, and tooling like an LSP, formatter or debugger support doesn't exist yet. The maintainer is actively recruiting help across tooling, standard library, backend fixes, and a security review of the TLS implementation.