September 19, 2026
Agent bugs never make it to the build: Bend rejects code without proof
Bend refuses to compile incorrect code: the agent rewrites until the compiler returns “All terms check.”.
Taelin
@victortaelin
Even in my most optimistic dreams, I did not expect anyone to grasp the essence the way Giulio did. And yes, that is the whole point. Models today are already incredibly capable; they have enough intelligence to build software of arbitrary complexity, but mistakes accumulate, and technical debt grows to the point where progress stops. If there were some magical way to tell an agent, “hey, this is wrong, try again,” until no bugs remained, the complexity of the software AI builds would stop hitting a ceiling altogether. Bend does exactly that!! Yes, the same task takes more time (and tokens), but that is precisely because writing correct code takes time! Bend accepts no incorrect output, and that is exactly what you want. Merged bugs create an illusion of progress, then bite you. I personally do not use other languages anymore. If Bend turns out to be a complete failure, it will at least remain the language I write my projects in. (and Bend written in Bend is already in the works)
for some reason, many smart folks seems to be kinda sleeping on Bend (??). so I want to give my take as to why this is so good. Bend 2 is basically a software factory language. the issue with models is that they will fuck up for your stuff without telling you, and in general cannot be trusted blind, thus making the "software factory" not super viable in practice. Bend acts as basically a replacement for what a .md does in a software factory. it defines the objectives and then it is up to the model to figure out the impl. you dont need to look at the impl because it is formally verified by the end of it. It's the whole point. YOU DONT NEED TO LOOK AT THE FINAL IMPL ANYMORE, THIS IS THE IMPORTANT PART. this is also why it is superior to lean. you can write the shittiest laws known to man in the most shitty and inneficient way possible and the final code will still be as performant as it can get within Bend's bounds. while disregarding concurrency and all that other annoying stuff.
· 193.1K views
Victor Taelin has settled the language question for himself: he writes his own projects only in Bend.
Syntax is not the point. Models can already handle software of arbitrary complexity, but errors accumulate into technical debt that brings work to a halt. Bend puts a filter at the agent's output, and the language simply rejects incorrect code.
How it works. A person writes laws in a LAWS.bend file: what the program must do. AI writes the implementation and proofs in PROOF.bend, and the command `bend PROOF.bend` verifies them. If even one law is unproven, the build fails. When everything adds up, the terminal displays “All terms check.”.
The spec used to be an md file. An agent would read an instruction, report “done,” and the person had to use their eyes and tests to find out whether the result matched the intent. Now the compiler verifies the spec: laws describe the goal, while the model figures out how to achieve it.
You pay in time. The same task consumes more time and tokens, and Taelin sees that as right: correct code takes time to write. The check itself is fast: Bend verifies 3 200 type instantiations in 0,38 s (explainx.ai review, 18.09.2026).
How to install it. Bend 2 was released on 17.09.2026; in two days the team shipped 16 patch versions. Installation takes one line: `curl -fsSL https://bend-lang.com/install.sh | sh`, then run a file with `bend hello.bend`, while `bend guide` opens the full guide. Linux or macOS is required; on Windows, the language runs through WSL. The language is free and open source.
A loosely written law will be followed to the letter by the agent: in third-party tests, it changed a character's movement and removed an obstructing flag. There is almost no standard library, so basic things have to be proved anew.
The Bend compiler is already being rewritten in Bend itself.
