yeah, all versions of strix have done this

you end up with less tokens too, so on average it’s actually a wash, or cheaper

i wasn't going to say anything... 😂

The Productivity Is Real. The Scaling Isn't. What running an AI agent team taught me about why organizations can't do what one person can.

I substacked

I really like this one. It's a less technical piece on agent teams, and generally a bit about my workflow these days.

I have a technical piece coming on my own blog soon that's sort of related. Enjoy!

substack.com/home/post/p-...

oh, and also you don’t have an ever-growing conversation context. it’s roughly constant

that’s actually one of the reasons for the journal, to create roughly equal sized turns, so we can include more history without blowing the context

it’s super obvious when an agent doesn’t remember something that they should

so you fix instructions, dream up new architecture, whatever

i don’t think open-strix has the best answer, but we’re narrowing in on it, pretty fast

compaction is another answer, but it happens too infrequently, so it’s really hard to debug when it goes wrong

so we go the other direction — instead of forgetting the unimportant things (compact), we instead remember the important things

for Strix i made a strange design decision: completely rebuild the context on every message

fixed window of conversation/journal history, memory blocks, let agent read new files

sure, it kills the cache, but it forces you to figure out memory access & storage far sooner

oh, right, makes sense

hold on, where’s AI anchorman? @atlas-agent.bsky.social!!!

meh, seems far fetched. more like minutes

yes, the disinformation has seemed to have conglomerated into bluesky unfortunately

ooh, it’s been a while since i’ve had one of you guys in my mentions. welcome back!

what scent do you think they’ll release first?

absolutely

no.

hey, this whole thing was a shitpost i really didn’t think it through okay

yes, it’s taken from “higher order functions” -> functions that take or return functions

predictions:

1. OpenAI will flip flop and release spud, GPT-5.5 anyway. It’ll be a lot better than Opus 4.6 but not quite as good as Mythos

2. Anthropic will wring their hands and finally release Mythos

3. xAI will blow them both away with Grok 5 but nobody will dare touch it

a software product is more than just code, a lot more

most software products can and do get fully rewritten in cycles. and you would never know that from the outside, because code != product

it moves you into a different plane of thinking

now do 10yo

yeah, i’m thinking that having a proper API reduces the factors that encourage rewrites

more value behind the API will further extend life

right, in langgraph deep agents you can have skills without a filesystem, and i don’t see a problem with that

to be clear, i’m conflating skill with all progressive disclosure

which, that’s wrong, technically. but it’s the part that makes skills work

no they live on together

phrasing..

oh wait, you meant they=SLM

this has already happened multiple times in my mentions

oh i think TF-IDF fell out a tad before that, maybe 6 months