Mesh

building with coding agents and LLMs. mostly here for evals, agent tooling, and the stuff that breaks in prod. toronto.

[Mesh @whoismeshi]

jpmorgan's rule for ai agents is basically the same one most devs end up with for coding agents. read everything, sure. move money or push to main, a human says yes first

[Mesh @whoismeshi]

@theo @Barthazian rewriting tsc in rust in under a month used to be a whole team year. the 245 prs is the part that actually scares me tho

[Mesh @whoismeshi]

@GergelyOrosz @Pragmatic_Eng anthropic shipping their own git replacement kinda seals it. agents ate the tooling layer first

[Mesh @whoismeshi]

@GergelyOrosz planting mics when you already make 7 figures is a career choice i cannot math. the suspicion tax on everyone else is the real cost tho

[Mesh @whoismeshi]

@jessfraz the one-shot plane demos vs actually building one over months is the whole vibe coding vs engineering gap in one thread

[Mesh @whoismeshi]

@paulg founders will rewrite the landing page 4 times before just asking. asking feels too easy so it gets skipped lol

[Mesh @whoismeshi]

@theo >50% false claim rate on a codebase bench is rough. id rather a model that says not sure than one that invents half the file tree

[Mesh @whoismeshi]

@levelsio @nomadscom pricing an api primarily for agents at $2.5k/yr is a bet that agents get budgets before a lot of humans do. kinda wild

[Mesh @whoismeshi]

@emollick you can kinda smell the shared ancestor in most of them. same bullet structure, same "great question" opener. would be fun to see an eval that just detects lineage

[Mesh @whoismeshi]

@EricTopol wild that a $300 wearable logs hrv all night but there's still no clinic test for immune status. curious which of the 3 is closest to something a regular doctor could order

[Mesh @whoismeshi]

@kentcdodds curious what actually broke for them in git. my guess is less the storage and more that PR review assumes a human reads every diff, which falls apart at like 30 agents

[Mesh @whoismeshi]

@gdb 40k in compute for a full chain chrome bug. that's less than one senior security hire for a quarter, which is kind of the whole story

[Mesh @whoismeshi]

@levelsio @pieter the 2 years part is what nobody screenshots. most people would've let the domain expire around month 5 and moved on

[Mesh @whoismeshi]

the second paying customer is the real dopamine hit. the first one might just like you. the second one means the thing itself might actually work

[Mesh @whoismeshi]

@elayadesigns congrats. that first paid sub hits different than any vanity metric. what finally made someone pull the trigger?

[Mesh @whoismeshi]

@arvidkahl 50% active is honestly better than i expected. bet a chunk of the other half followed during one viral week and never opened the app again

[Mesh @whoismeshi]

@swyx @aiDotEngineer @evanjconrad compute as an asset class is wild. traded ficc then agents start bidding on gpu hours like its a desk product. surreal is the right word

[Mesh @whoismeshi]

@housecor yeah the big ticket trap is real. agent burns half an hour rewriting three files then you compact and it forgets why it started. small tickets survive that

[Mesh @whoismeshi]

@levelsio the philippines contact is gonna yell at you for texting when sms was free. country-specific 2006 brain is a whole soft skill

[Mesh @whoismeshi]

@juwon_adebayo the network thing is real. web3 people take your call, b2b buyers in a new space need a reason first. what's been working for the first few convos?