I recently joined @replohq, the easiest way to do Ecommerce on the internet.
Two weeks ago I built an internal AI agent called Replee in my apartment.
It’s now used across every part of the company, and has made 326 code contributions as of today.
What started as a way to better understand OpenCode and triage Sentry tickets has turned into general purpose background infrastructure for the Replo team.
Below are some ways the Replo team uses Replee today, and what your team’s own version could do too!
(note: shout out to @rahulgs, who inspired me to start working on Replee with this tweet: https://t.co/Ma1fzdOifF)
we built our own background coding agent at ramp: it's called inspect
its powered by @opencode and @modal, and works with all the frontier models
it has a cloud hosted version of vscode, chromium, and terminal. has all the tooling and skills a ramp engineer would have, and is blazing fast
we're basically giving every ramp builder infinite laptops so they can yolo their most ambitious ideas at zero marginal cost with the biggest models we can find
@zachbruggeman@monasticpanic@nzgb
at first glance, this is really dumb
but a potential positive consequence is that there will be way more time and energy spent on better LLM inference that any company can use, which is the real barrier at this point between using Claude Opus vs GLM/Kimi for production use cases
for an example this past week I used GLM 5.2 and it was ~GPT 5.5 quality, but the inference providers I used were subpar relative to OpenAI
New w/ @leomschwartz@amir:
The Trump admin has asked OpenAI to stagger the release of GPT-5.6 over security concerns.
On Thursday, CEO Sam Altman told staff that the government will be approving access to GPT-5.6 customer by customer, a highly unusual approach.
One of the things I’ve hated using other apps is terrible customer support or being passed off to a bot - especially when money is involved.
With Hype we try to make sure you have the absolute best support and extremely fast response times where you can talk to a real human ASAP
I’ve been using exclusively 5.5 on low thinking for the past month ish, my bigger thing is speed
I feel like the bottleneck is how fast an llm gets something done well which is why I have done this
Ive been working with 5.5 on medium for planning and then 5.4 mini or Kimi k2.6 for implementation and it's been fantastic. People need to stop using 5.5 or opus for everything
Company Brain
@t_blom
Every company has critical know-how scattered across people's heads, old Slack threads, support tickets, and databases, and AI agents can't operate like that.
We think every company in the world is going to need a new primitive: a living map of how the company works that turns its own artifacts into an executable skills file for AI.
Due to the unmatched commerce capabilities of @replohq Agent, we have delayed launching it to the public.
We gave it access to a Dropbox of a store we were thinking about building, and we found out Agent built it when one of our engineers got served an ad for the store while eating a sandwich in the park.
A Stanford study found that people who played Pokémon heavily as kids developed a small region of the brain that responds specifically to Pokémon characters.
Researchers scanned adults who grew up playing on Game Boy and showed them images of Pokémon like Pikachu and Bulbasaur.
Their brains lit up in the same exact spot, a consistent area in the visual cortex tied to recognizing specific categories of objects.
The reason comes down to childhood. When you’re young, your brain is more flexible, and spending hours memorizing hundreds of similar-looking Pokémon essentially trained it to carve out space just for them.
(via @Stanford)