Agree with the sentiment. To me things seem to be interoperable more than ever before even with agents so maybe not fragmented in the traditional sense but there are so many options that would work for any workflow. People are trying to figure out what works best right now.
Maybe custom setups are just the new normal going forward. I personally believe when lots of startups are doing the same things slightly differently, it means there's a central idea they are revolving around that someone will crack eventually.
I've been working on that problem a lot recently. Both from a product review and code review standpoint. And I think code review as it was will never work anymore.
Product validation loops are everything and building those validations autonomously is the next stage I'm working on.
Ideally it should be the agents job to show you that their output is working.
It's interesting to see the wide spectrum of software engineering in companies today.
The most AI pilled people have built layers of coordinating agents to build full apps nearly hands free while some people are still just getting by with just claude code or cursor as an assist.
Why aren't more people going all of the way? It at least seems to be functional with what we have seen built.
Is the jump too difficult? Is it too expensive?
@tomhacks@SergioTapiaDev I was just building a tool to help with this for my own development. A visual mental map over a code review with blast radius calculations built in to speed things up. Still just starting https://t.co/cCOPjKhzB9
Experimenting with visually mapped PRs with precomputed blast radiuses to allow for more efficient code reviews. Not sure if it has legs but it was cool to at least dog food it on its own development here.
Very rough but let me know what you think!
I posted a screen cap here. It makes the review more visually organized at least if you are familiar with the codebase but it's rough enough where I don't know if the things I don't like about it are structural or just aesthetic. Will have to keep using it to see I guess.
https://t.co/esYehJymQb
Experimenting with visually mapped PRs with precomputed blast radiuses to allow for more efficient code reviews. Not sure if it has legs but it was cool to at least dog food it on its own development here.
Very rough but let me know what you think!
Works by turning the code into a graph and mapping that graph into an architecture diagram (can be user or agent made using MCP) and then mapping that to risk by giving it criteria like hard files to mark critical or abstract statements like "directly impacts user". The agent actually notifies if the map changes on each PR to keep it up to date as well.
Was trying to accomplish 2 things, focus on getting the gist of a PR rather than reading every line while remaining safe and keeping a mental map updated at all times so I could still understand the code as it scales.
Experimenting with visually mapped PRs with precomputed blast radiuses to allow for more efficient code reviews. Not sure if it has legs but it was cool to at least dog food it on its own development here.
Very rough but let me know what you think!
@TPateeq@boatdotdev@Agent37HQ I'm playing around with a new project, do you have one you would recommend in particular? I haven't tried any of them yet. I'm not doing anything of scale yet tho so not sure it even matters.
@neilkale@rikiyatakehi@jayden_teoh_ Do we still have any intuition why models are able to do weight space programming at all? Are they simulating samples and tweaking based on activations they see on the fly or do they have some implicit understanding of how the weights should be?
@sachintwtss Maybe. I'm just seeing as the models are getting better and better, maybe the job has changed again. I'm getting close to a new possible approach to review, I'll be happy to share some snippets (will not be pretty) once i get something tangible