Leonard Bedner

Head of AI Engineering at Betr Health.

Building Aegis Stack. 25 years shipping. Code structure is more important than the model.

uvx aegis-stack init mvp

[Leonard Bedner @leonardbedner]

@badlogicgames lol

[Leonard Bedner @leonardbedner]

@valigo If only.

[Leonard Bedner @leonardbedner]

@MaritvanDijk77 100% agree. I have an app that does a lot of calculations, and it's done deterministically ahead of time before the model even touches it, so the model can just worry about telling me about it versus trying to calculate it for me.

[Leonard Bedner @leonardbedner]

@boristane Did you... tell it to? My agents are constantly refactoring and removing dead code.

[Leonard Bedner @leonardbedner]

Nope, because caring deeply about the code and patterns is what allows my agents to move fast through my code. Highly structured, follow patterns I have used for years, etc. None of that has gone away. It's empowered me, if anything.

In my experience, anyone say that is either not experienced enough yet, or they're just parroting what other say.

[Leonard Bedner @leonardbedner]

Added granian as an officially supported ASGI server alongside uvicorn for Aegis Stack.

Uvicorn stays the default, benchmarks below.

Next up: a deeper look at zuvloop (already available for uvicorn on Python 3.14) and zttp, the Zig replacements for uvicorn's two speed dependencies.

[Leonard Bedner @leonardbedner]

@ckathanasiadis @michael_timbs What magic is this?! *Goes off to check pytest*

[Leonard Bedner @leonardbedner]

@michael_timbs I have been thinking about this for a few years... Every time it comes up, it's always the "Just because you have 100% coverage, it doesn't mean you don't have apps", and that slowly moves to "Unit tests aren't good".

I just assume they never saw good test infrastructure before.

[Leonard Bedner @leonardbedner]

@earthlingworks For me, it's more about having control over the software that I have a feeling is going to go to shit over the next few years. I have no faith in these companies lately.

Plus, depending on what the software is, my solution may end up being better.

[Leonard Bedner @leonardbedner]

@YashHustle_22 Because 90% of that code big companies have generated are slop grenades, things skilled engineers have to first clean up, then make sure never happens again.

[Leonard Bedner @leonardbedner]

@Ritesh_0_9 @YashHustle_22 I would respond that you mostly have it backwards. Skilled engineers can orchestrate agents like a symphony, including in production.

Beginners are the worst with AI, because they don't have enough experience to properly drive it.

[Flet @fletdev]

Flet 1.0 is here! 🎉

Build web, desktop, and mobile apps in Python.

Ready for production apps, with declarative UI, Python libraries on mobile, app testing, and AI tools in Studio + MCP.

Four years in the making. Thank you, Flet community! ❤️

https://t.co/dNBnRxYs6J

[Leonard Bedner @leonardbedner]

@AtMemAi I rolled my own memory system, and then just inject it into PydanticAI's ctx as/where needed.

[Leonard Bedner @leonardbedner]

Well okay then... I guess I'll just stay right where I am?

[Leonard Bedner @leonardbedner]

@unclebobmartin I would give @pidotdev a look. As minimalist as it gets.

[Leonard Bedner @leonardbedner]

@mateusz_palak I'd argue Rails overall as well. Drop an agent in there, I'm sure it has enough training data, plus literally the conventions in front of it, it should do well.

[Leonard Bedner @leonardbedner]

@unclebobmartin So we are gods?! Nice! 😅

[Leonard Bedner @leonardbedner]

@unclebobmartin That's what I'm building, so I hope so 😅

[Leonard Bedner @leonardbedner]

@mischavdburg This is what I always tell people when they want to learn XYX. Build a full app that uses XYX, something you would use and care about. Then keep building, keep iterating.