the most underrated thing an agent can do is abstain
every demo shows what the agent did. nobody demos what it refused to touch
a confident wrong answer costs more than a slow honest one
@mirku21 the swarm split makes sense, cheap models for bulk and the expensive one for judgment
the part that never makes the demo is who reads what the swarm merged
@cryptosoha66 an agent putting software on the company bill needs approval scoped to the exact item and price, not a budget it interprets
'the agent thought it was allowed' is going to be a fun incident review
we launch lear on product hunt on tuesday
the feature i care most about doesn't demo well. it asks before it runs
turns out trust is hard to screenshot
@Evan_Lin an open source ai sre hooked into grafana and loki is the right shape for this
what happens when it thinks it found the fix, does a human approve before anything runs?
someone drained $4.86m from chainflip yesterday in one transaction
nine minutes later the eth was already in tornado cash
worth asking whether your own incident plan moves in minutes or in meetings
@_mohansolo does swe-sweep grade the fix or just the find? catching the bug is the easy half
a fix that quietly breaks a caller three files away is where agents get expensive
@undefinedKi the mechanism rings true even if 90% is a big number. reading whole files to find the twenty lines you need was always the wasteful part, whatever model sits on top
@gagu_sh the agent writing the test that kills its own surviving mutant is a tidy loop
do you check the new test actually fails on the mutant, or trust it because it passes on the fixed code?
@arosplatforms labelled partial until a person signs off is the right default
does the owner see the abstained steps inline with the answer, or is the sign off happening somewhere else?
lear runs inside your infrastructure
i'd rather explain that plainly than make it sound like you hand us production and hope for the best
it proposes the fix. you decide whether it runs
@cammcbride@thinkymachines@dok2001@zatlyn@michellechen the kba matching is the part i'd expect a small model to be weirdly good at, it's mostly retrieval over past symptoms
how are you checking better there, a set of past hosts where you know the right kba?
"we're investigating" is fine for the first update
after that i'd rather read which service you're checking and what you've ruled out. even if the answer is still that you don't know