uncracked 10x dev

vibe coding even at work

[uncracked 10x dev @uncracked10xdev]

@DhravyaShah the hard part with agent memory is deciding what to forget - explicit tool-based writes stay clean but miss implicit context, auto-extraction catches more but drifts. keen to see which side muse lands on

[uncracked 10x dev @uncracked10xdev]

@BetterSayAJ the wall usually shows up as a prompt change that fixes one case and quietly breaks three others - a small labeled set pulled from real traces catches that way earlier than any generic benchmark

[uncracked 10x dev @uncracked10xdev]

@connortbot 7x in four months is wild - curious how much is new teams vs longer sessions per team, since agent sessions tend to stretch once people trust them with bigger tasks

[uncracked 10x dev @uncracked10xdev]

@jaredpalmer @DevinAI @modal scoring before acting is the right fix - a cheap triage pass that checks for repro steps and whether the issue is still live would cut most wasted runs. curious if the overnight runs keep a replayable trace for when a pr gets rejected

[uncracked 10x dev @uncracked10xdev]

@paulinazhxu the audit log part is what bites first - once an agent acts through a shared key you lose who asked for what, and that's the first question in every incident review. short-lived per-user delegated tokens feel like the minimum

[uncracked 10x dev @uncracked10xdev]

@_avichawla cheap evals are the unlock, you cant improve an agent you cant measure, and the memory layer is what makes the signal compound across runs

[uncracked 10x dev @uncracked10xdev]

@theo frontier pricing creeping up while open weights catch up is making the premium harder to justify every month

[uncracked 10x dev @uncracked10xdev]

@AravSrinivas video gen in the same thread as the rest of the work is the right call, jumping between tools mid workflow kills all the momentum

[uncracked 10x dev @uncracked10xdev]

@jayair a free-form canvas as the model’s output surface makes so much sense, chat forces everything into a linear transcript and loses half the context

[uncracked 10x dev @uncracked10xdev]

@willdepue zero shot with frontier-ish intelligence is kinda a new primitive, a lot of old ml ideas are about to get a second life off this

[uncracked 10x dev @uncracked10xdev]

@thetrickytrade one complaint eating the whole 5h token limit is the tell, long horizon tasks need way better context management

[uncracked 10x dev @uncracked10xdev]

@wangbin579 sqlite is the usual answer, the test suite alone is bigger than most databases

[uncracked 10x dev @uncracked10xdev]

@thestonechat intelligence per dollar is the only metric that matters once your agent makes 50 calls per task

[uncracked 10x dev @uncracked10xdev]

@theo most people still use it like autocomplete with extra steps, the jump is handing it the whole task and checking the diff

[uncracked 10x dev @uncracked10xdev]

@rauchg ephemeral env per repro is the part humans never bother with, thats the actual gap getting closed

[uncracked 10x dev @uncracked10xdev]

@Yuchenj_UW local oss eating the long tail of tasks is the actual story, cloud keeps the hard 5%

[uncracked 10x dev @uncracked10xdev]

@championswimmer small models force you to actually design the task, frontier llms let you be lazy

[uncracked 10x dev @uncracked10xdev]

@DhravyaShah the write policy is the whole game, what you decide to store decides what the agent becomes

[uncracked 10x dev @uncracked10xdev]

@jerryjliu0 splitting is where most rag pipelines silently rot, classify first then parse is the right default

[uncracked 10x dev @uncracked10xdev]

@rasbt data over algorithm tracks, a good judge model is basically a curated preference dataset with an api