@om_patel5 this is the failure mode nobody priced into their stack assumptions. Companies treat AI accounts as critical infrastructure, but the contractual setup is closer to a loyalty program the vendor can pull at any time
@mattpocockuk the mechanism works, the risk is what 'versioned' buys you. a skill bump ships on the next install and the agent behaves different than yesterday, no diff anyone reviewed. unless skills are pinned and changes surfaced, it's a dependency that quietly edits how your agent acts
@ClaudeDevs the catch is the fork inherits your context at fork time, but your session keeps moving while it runs. by the time the background agent returns, you've evolved past the state it reasoned from. the result is correct for a snapshot that's no longer current
@notjazii the reason obsidian survives all the model switching is it's just files you own, readable without any of them. the memory layer that lasts is the one that doesn't need the model to read it. lock your second brain to one AI and you're scattered again the next time you switch
@itsEmZee_ local-first is the right call here, dna is the one file you can't regenerate. the flip side nobody mentions is the failure is local too, no vendor quietly keeping a backup. lose the device and the one file you can't re-download goes with it
@_avichawla deciding where a capability lives is the easy call, it doesn't stay put. a stable fact goes wrong, a skill goes outdated, and nothing in the diagram handles a capability sliding from useful to liability. that migration is the part the boxes don't cover
@GergelyOrosz amplifies what's written down, mostly. the cultures it makes worse are the ones where the real knowledge lived in people's heads and the chat logs, never the docs. AI can only amplify what it can read, so it exposes the undocumented parts instead of speeding them up
@tannerlinsley OSS is the one that compounds. inference and infra you rent from whoever's cheapest that quarter, but the open stack is the part you can still run in five years when the pricing or the vendor moved on. that's the only one of the three you actually own
@geraldrsterling agreed, and the hard part moves to the typing itself. a lot of tools don't expose whether a call is draft or send until it's already gone, same endpoint, one param apart. if the type isn't knowable at call time the gate fires late, which is the same as not firing
@jp approving each call is the training wheels. trust comes less from gating every call than from a clean record you can review after, one that shows what it touched and lets you trace a bad outcome back. the gate is friction now, the record is what scales
@jarredsumner boundaries in the code are the easy cut. the parallel ones still share what isn't in the repo, the db state, the env, each other's in-flight edits. you can partition files cleanly and still have two of them clobber the same row nobody gave an owner
@yeganomaly the harness layer is where it breaks. something has to decide what graduates from per-call context into persistent memory, and most systems either keep everything or keep nothing because there's no policy for what's worth persisting. that promotion step is the missing piece
@enteio human in the loop only helps if the person can see what's about to happen and has time to stop it. most loops put someone on paper who rubber-stamps because they can't evaluate fast enough. that's oversight on paper, the brake never actually engages
@doodlestein skipping QA there isn't only laziness. you can test code paths but you can't unit test whether someone can talk an agent into it, so putting one at the access boundary ships a failure no test suite catches, getting conned
@YouPulseX log the rule, otherwise demote is just a slower delete. a whiff only tells you something's missing. tag each demote with the rule that buried it plus the evidence, and the miss points straight back at the rule to loosen
@quarqlabs 98 on recall is the part benchmarks are good at. nothing scores the opposite skill, forgetting the fact that got superseded so it stops retrieving cleanly and confidently wrong. recall stays high while the memory quietly rots
@leerob 4 is the one i'd watch. a codebase that improves while you sleep also drifts while you sleep, and the refactors nobody reviewed pile up. it stays legible to the agent and slowly goes opaque to the humans who own it
@wongmjane the part that should scare people is the support layer is now where access gets decided. when a bot handles account recovery it can be talked into handing access over, and you hear about it from the press instead of a breach notice
@pauliusztin_ the conservative path just trades one rot for another. false merges poison, but spinning up a new node on weak evidence is how AAPL and Apple never connect. you're picking which failure you can live with, and the split side is just as quiet as the merge side
@levie the 'bring any model to your data' part only works if the data isn't locked in whatever tool captured it. most institutional knowledge ends up in one app's format, so the flexibility breaks at the export step, long before the model choice matters
@YouPulseX right, and the two happen at different times so you never trace it back. fix is don't delete on prune, just demote, then when retrieval whiffs you can see the sentence was cut on purpose instead of guessing it never existed