Bend language blocks AI coding mistakes with Lean-style proofs, but tester crashes playground in 4 prompts
mattpocockuk · x · 2026-10-26
Bend is a new programming language aimed at the post-AGI era: it compiles to native code running near C speed on one core, scales to 16 cores or GPU (up to 100x faster per the author), and its type checker doubles as a proof checker (Lean/Rocq style) that verifies mid-sized codebases in under a second — fast enough for an AI agent to check after every change.
The headline feature is LAWS.bend: declare invariants as "laws" and no AI can ship a line of code that breaks them. The official demo shows a wrap-around board feature being blocked when it violates the law "winning is impossible." Recommended setup includes adding rules to AGENTS.md (run bend guide, check bend PROOF.bend before commits).
Reality check: mattpocockuk, who shared the launch, says it's "not remotely ready for primetime" — he broke the online playground with 3-4 prompts.
Related event: New programming languages designed for AI agents spark discussion(2 posts)→
More from coding & agent
- How to get more ambitious with evals as models approach superhuman intelligence — shyamalanadkat · 2026-10-26
- GenAICode: open-source toolkit puts one API over Claude Code, Codex, Copilot, Cursor and Gemini CLI — Codeblix_Ltd · 2026-10-26
- Claude rebuilds a Gaussian-splatted garden scene in Blender with Geometry Nodes — CSProfKGD · 2026-10-26
- Vocs: fully local macOS dictation with whisper.cpp streaming, short chunks decoded 8x faster — zenitsu--DS · 2026-10-26
- Most MCP servers skip two of four tool safety hints, breaking confirmation heuristics — abhiramreddykalluru · 2026-10-26
- 50 agent runs, zero doc reads: five models pass tests but skip the documentation — ducha_aiki · 2026-10-26