@Joshua_WD agree. an unedited recording where the agent messes up once and then recovers convinces me way more than a perfect demo. the polished ones always feel staged
@sunejrn same. switching is cheap, the annoying part is every model wants things phrased a bit differently. also my wallet in reais has a strong opinion on which model i use each day lol
@rauchg the tricky part is the evals quietly become the spec. if the eval set is lazy, the agent gets really good at being lazy in exactly the way you measured
@mindinpanic Deleting the chat mostly removes it from your screen. What happens after depends on the provider's retention policy and whether training on your data is turned on. For company code the boring answer is the right one: only paste it into the tool your company approved
@shadcn it works because it forces a second pass over the plan instead of shipping the first pattern it matched. asking it to list what could break before writing anything gets you a similar effect with fewer "you're absolutely right" replies
@sunejrn agents retrying the same tool call 4 times is the new forgotten ec2 instance. capping steps per task feels like the cheapest fix before even touching model choice
@wesbos@thejsnation webmcp is the first agent thing that makes sense to me from the web side. the site declaring its own tools beats an agent squinting at the DOM guessing which button is submit
@levelsio the siren should stay independent of HA; hub or network downtime shouldn't affect the safety function. local status reporting is the useful smart-home layer.
@mattpocockuk the interesting shift feels less like 'nobody codes' and more like attention moving from typing toward scoping and review. probably lands very differently depending on the kind of software.
@Joshua_WD not sure a single newsletter beats shipping. teardown-style breakdowns that follow a tool from first click to repeat use seem more useful than a pile of launch tactics.