David

introspecting about agents, economics and health – vibes are my own

[David @dasfacc]

2026 is the year of accountability

[Agustin Lebron @AgustinLebron3]

This blog post still lives rent-free in my head.
https://t.co/0igY7hy2uK

I think the AI labs would do well to read it.

[David @dasfacc]

@MParakhin vis-a-vis*

[David @dasfacc]

@paularambles what the foo bar you talking about

[David @dasfacc]

am i stupid?

I must be because I read people on here complaining about @OpenAIDevs codex bugs draining system resources

then I open codex later in the day and reactively click the update button

just for codex to almost crash me at 15 GB of RAM

did you guys move compaction client-side or something?

[David @dasfacc]

devil's advocate here but given that we don't have a good idea what's easy and what's hard for these models it may very well be that curing cancer is regular business for them whereas the illicit distillation sniffing is a pain in the ass

specially bc you can use ai to get creative at illicit distillation, cancer can't (thank God)

[David @dasfacc]

@BenceRedmond we found greptil unusable (4 months ago?) but devin convinced us and we've used it ever since

additional to our local review setup though

capy sounds interesting!

[David @dasfacc]

@itsericlay what is this bar

[David @dasfacc]

@enggirlfriend do you prefer ... ? what is there to prefer lol

[David @dasfacc]

I never user low or mid, im always on high and go extra for the new/complicated stuff

I've a friend that's always on mid and he's happy

if you want to determine a baseline you could identify repeatable workflows in past sessions

I wouldn't mind you A/B testing medium on me to determine sufficiency, but if it fails I definitely want a way to turn it off for a month, or sth like that

[David @dasfacc]

@hyojun_at there's a funny quirk where trying to drag a tab into the bookmarks instead moves drags the window. happens only when there's only one tab open lmao

trying it out :)

[David @dasfacc]

@Rasmic holy fuck

[David @dasfacc]

1. have a manager worktree
2. n number of managers running off that worktree
3. disallowed to do code stuff, must use subagents for everything
4. tell it to adopt any issue or PR
5. creates worktree, opens in tmux session `agent-n`, runs coding agent
6. tmux agent does its thing, gets it green and reviewed
7. manager enforces rules, finetunes agent CEs and envs, creates learnings, clears context when too large, tears down agents/worktrees after merge

[David @dasfacc]

@migueldeicaza I can confirm

[David @dasfacc]

@PatrickToulme the apple of inference? interesting thesis

[David @dasfacc]

@ben_ai_eng I don't work at the labs but you've my follow!

[David @dasfacc]

@cramforce filled crust region

[David @dasfacc]

this is my own fault/btw/for using/so/many/slashes

[David @dasfacc]

yeah please proceed with the... uuhh

[David @dasfacc]

v2 is out of beta! capable as ever

[Pydantic @pydantic]

𝗣𝘆𝗱𝗮𝗻𝘁𝗶𝗰 𝗔𝗜 𝘃𝟮 𝗶𝘀 𝗵𝗲𝗿𝗲, and your agents have never been more capable. The inner loop of an agent is settled, the leverage is in the layer around it, and v2 turns that whole layer into one thing you compose: the capability.

It bundles instructions, tools, hooks, and model settings into a single unit. Plus a leaner core and the Pydantic AI Harness.

[David @dasfacc]

@paularambles that’s confusing ah

is it un taco? or dos tacos?

is it uno mas dos igual a tres tacos? 😵‍💫