Bend Introduces a Proof-Oriented Parallel Language for AI-Coded Apps

Summary

Bend is a programming language designed for applications built with AI coding agents. Its central approach is to make human intent more precise through executable laws and to require proofs that generated implementations preserve those laws. The language uses Python-like syntax, compiles to native code, and is presented as running nearly as fast as C on one core. The same compiled program can run across multiple CPU cores or on a GPU, with the project claiming speedups of up to 100 times over one core in suitable parallel workloads. Bend's runtime automatically spreads split work across available cores and joins the results, so developers do not need to write threads, locks, or GPU kernels. Its type checker also acts as a proof checker based on ideas associated with Lean and Rocq, while the project claims checks can finish in at most about a second on its target workflow. Developers can place invariants in a `LAWS.bend` file and have an AI agent implement corresponding proofs in `PROOF.bend`; the supplied game example uses this mechanism to block a change that would violate a declared rule. The site recommends adding Bend-specific instructions to `AGENTS.md`, including running `bend guide` and checking proofs before commits. Bend is still evolving and may contain bugs; the project says it works best on back-end projects running on Linux or macOS. The language is supported by the BendTT type-theory paper and BendRT parallel-runtime paper.`