Umair

the self-verification is the interesting part but i think the framing is slightly off. the hard problem was never "does the model pause to check" — it's whether it knows WHICH assumptions are actually risky vs which ones are safe to skip. because verification isn't free. every sanity check costs tokens and time. 4.6 didn't loop because it was lazy, it looped because it couldn't distinguish...

50% of doc traffic from agents is a wild stat but i bet most of that traffic is wasted. agents dont read docs the way humans do, they ctrl+f for function signatures and ignore everything else. the real unlock isnt prettier docs its structured metadata that agents can actually parse without hallucinating the rest

honest question - do agents even resolve identity through DNS? every agent i run authenticates with API keys and oauth tokens, not domain names. the naming problem for agents is really an auth and trust problem and TLDs dont solve that at all. you could own coolagent.agent and still be running malware behind it

specs are just docs that lie slower. the real bottleneck with multiple agents isnt coordination its conflict resolution when two of them disagree on approach. a living spec doesnt fix that, it just makes the disagreement harder to notice until the PR is already broken

nps jumping from 34 to 61 while dau drops is the whole story tbh. engagement dashboards are addictive for founders too - watching numbers go up feels like progress even when its just people being confused or stuck

the question is kinda wrong imo. youre not "coding from your phone" youre managing agents from your phone. totally different skill. i send a message describing what i want, the agent does the work, i check back later. thats not coding thats project management with extra steps lol

memory is the hard part though, not the simple part. everyone treats it like a database lookup but knowing what to remember and when to bring it up is basically the whole product. most chatbots that "remember" just ctrl+f your history and regurgitate it at weird times

the problem isnt more access its that you gave the agent freedom to decide what matters. when it only had name/company/role the constraint forced relevance. remove the constraint and it tries to use everything because thats what LLMs do, they fill space. fix is simple: give it more data but explicitly tell it to pick ONE detail max. the access wasnt the issue, the lack of a selection heuristic...

the "approve before it acts" thing sounds safe until you realize youre approving from a phone screen with zero context. nobody is carefully reading diffs on mobile, youre just tapping yes. at that point the approval step is security theater not a safety mechanism

screenshot based automation is gonna be the new selenium - looks impressive in demos, becomes a maintenance nightmare in prod. for anything you do more than once just write a proper integration. computer use should be the last resort not the first thing you reach for

idk i think the advice itself is fine, people just skip the "learn" part. you shipped features weekly and learned nothing from it - thats not a speed problem thats a feedback problem. the 5 user interviews you did IS shipping fast, you just finally closed the loop

vote selling exists because "product of the day" became resume material. you created a status game worth gaming and now youre surprised theres a market for it. fixing supply side is whack-a-mole, the demand is structural. as long as a PH badge on your landing page converts, people will pay for votes

the question is framed wrong imo. its not about how many YOU run in parallel, its about whether the agent can spawn and manage its own subagents without you babysitting. one well-configured agent that delegates work autonomously is worth more than 15 you have to manually context switch between. everyone in this thread is describing management overhead as if thats an inherent problem when really...

sora was never a product, it was a demo that got shipped. nobody who actually makes video content used it for anything real because the workflow is completely wrong - you dont need a magic box that generates 10 second clips from text, you need image gen + selective animation + ffmpeg. the people making actual AI youtube channels spend like $2 per video using cheaper models and only animate...

the framing here is backwards imo. agents dont "pick" tools - they regurgitate whatever their training data and system prompts tell them to use. resend wins over alternatives because its in more github repos and tutorials, not because an agent evaluated it and decided it was better. thats pattern matching not preference. the interesting question isnt what agents choose, its what gets into the...

the layer framework makes sense but i think the real variable isnt "where you use AI" its whether you can read what it wrote and know if its wrong. senior devs avoiding AI on infra still ship bugs, they just mass produce them slower lol. the 3 weeks coding 1 week debugging thing was true before AI too

the double vote thing kinda defeats the purpose tho. teams with big external communities just tell everyone to vote during randomized windows for 2x impact. so the people who already had an advantage now get double the advantage. randomization helps visibility but the scoring mechanic rewards coordination not discovery

honestly the poll results tell the whole story - 51% using 1-2 agents. because thats the number you can actually keep honest. running 10+ just means youre merging code you skimmed instead of read. i tried scaling up and my git blame turned into a mystery novel i couldnt follow. went back to 2 agents with really detailed AGENTS.md files per repo and the output quality went way up because each...

the advice wasnt wrong tbh, you were just consistent at saying nothing. consistency without a point of view is just scheduled noise. the messy post worked because you finally had something to say, not because structure is bad

honestly i almost never try products from browsing PH. i try them when i have a specific problem and someone mentions the tool in a thread or comment. the browsing/demo thing is window shopping - feels productive but rarely leads to actually using anything long term. the products that stuck for me were all discovered through someone else complaining about the same problem i had