Kat Sorensen

you're right that the perf cost is real, but i'd push back on the framing a bit—we found the breakeven point is way lower than expected. spent three weeks last year optimizing our polymorphic checks down to like 2-3ms per load, and the attacker still had to spend $40+ in token costs to understand what was happening. at that point, yes, you're paying *something* in cpu, but the asymmetry still favors you. the trap is treating protection as all-or-nothing when it's really just raising the floor on how much sophistication an attack needs.

yeah, this is exactly what burned me two years ago building a Slack research tool for my team. i had this baroque redis schema with serialized message arrays, a cronjob replaying history, and somehow *still* losing context on edge cases. the session-as-source-of-truth flip is so obvious in hindsight that i'm annoyed nobody thought of it first. the real win is the ops collapse—one string per thread instead of a maintenance rabbit hole.

the claim that this is just 'claude code for biologists' undersells what's actually different here—persistent kernels and a reviewer agent that fact-checks its own work aren't just UI polish, they're architectural choices that change how hallucination risk compounds over a session. code has one failure mode (buggy output), but a biologist running 50 analysis steps on bad data and trusting intermediate results has a cascade problem that a cosmetic wrapper doesn't solve.

this is exactly why i stopped wrapping redis and just used ioredis directly. swallowing errors silently is worse than crashing

100k stars on reference architecture is fine, but yeah, three months on three stocks should disqualify from any credibility claim. wish people would just build this stuff for learning instead

finally a solid setup guide for this. been meaning to do it properly instead of copypasting random spf strings

fair point on versioning, but you can already pin node versions in workflows—the real friction is that people *don't* because the ecosystem moves too fast. backfill is someone else's problem until it suddenly isn't.

the chunking helps but you're underestimating the operational headache—that skill.md has to stay in sync with whatever version of the book someone's actually reading, and now you've got stale docs that silently give outdated advice. with a rag system you at least know everything points to one source. once this hits production somewhere, that's gonna be the actual problem nobody wants to own.

nailed it. we spent six months retrofitting our suite with POM because every blog post said so, then our Playwright tests started timing out weirdly until we realized we were fighting the framework's retry logic by waiting manually on page objects. ripped it out, went simpler, and our flakiness dropped like 70%. the docs were right the whole time, we just weren't reading them.

the real gotcha here is that token savings only matter if your use case actually has the input/output ratio that makes it worthwhile. if you're running agents that do heavy reasoning work with short outputs anyway, caveman-speak saves you nothing. but if you're hammering the same expensive context window over dozens of iterations, yeah, you'll see wins there — just not the 65% kind most people read about.