With this, I'm closing the loop with three nice tools: Koubou, which automates screenshots for the app store; Mav, which enables the agent to drive the simulator and record evidence; and Musts, which puts everything together.
The agent records evidence. When musts is happy, the loop closes and the changes are done, for real. Please give it a try and let me know if it works for you.
https://github.com/bitomule/mustshttps://blog.bitomule.com/posts/introducing-musts
GitHub - bitomule/musts: The validation loop that stops AI coding agents from claiming work is done before it actually is.
The validation loop that stops AI coding agents from claiming work is done before it actually is. - bitomule/musts
So today I'm introducing musts, a small Rust CLI that creates a validation loop. The agent runs musts validate when it thinks it's done. musts knows what changed in your project and returns the validations that have to run before the task can close.
Some example cases: if swift code changed review using this skill. If a view was changed ensure this was used. If this module changed build only this changes.
Yesterday I shared mav. I have been thinking a lot about validation: making sure changes satisfy a given rule like tests passing.I tried everything but it still required manual intervention from me. That did not scale.
AI coding agents don’t fail because they write bad code.
They fail because they can ship “good enough” before you’ve done the thinking that would make you better.
I wrote about why learning to program still matters, even when an agent can do it for you:
https://blog.bitomule.com/posts/learning-to-code-when-the-agent-can-code-for-you/
Learning to code when the agent can code for you | Bitomule's learning shack
AI agents make code easier to produce, but not engineering easier to learn. The work that builds judgment has to be protected on purpose now.
Tiny UX detail, big user impact: Boxy now makes Family Sharing clearer before checkout. If you’re organizing a move with someone else, the last thing you need is subscription confusion on top of cardboard chaos 📦
Introducing mav.
It is a small CLI that gives coding agents a practical way to drive an iOS simulator or device, run repeatable flows, and create evidence a human can review.
https://blog.bitomule.com/posts/introducing-mav/
Going all in this week before Code with Claude Code in London on Tuesday.
AI coding agents don’t write bad code.
They write good-enough code before you’ve done the thinking that makes you better.
I wrote about why learning to program still matters:
https://blog.bitomule.com/posts/learning-to-code-when-the-agent-can-code-for-you/
Learning to code when the agent can code for you | Bitomule's learning shack
AI agents make code easier to produce, but not engineering easier to learn. The work that builds judgment has to be protected on purpose now.
Lately I am thinking a lot, and talking with many people, about how LLMs changed software engineering.
My current take: code is no longer the bottleneck. Validation is.
Trying to revive my blog for the 13421498th time to think in public:
https://blog.bitomule.com/posts/code-is-no-longer-the-bottleneck/
Code is no longer the bottleneck | Bitomule's learning shack
AI agents did not remove the bottleneck. They moved it from writing code to validating what is worth shipping.
