In the linked story, the developer using the agent went wrong when they allowed the agent to have the autonomy to handle a real life social interaction that requires very careful judgment. Was the agent instructed to read the comment thread and handle the issue arbitrarily? Not gonna go well.
I suppose you could just tell it to approach the situation unbiased and lean towards lateral thinking and considering all options for the best outcome. It will always be biased in some way as it has to choose something. That’s its goal is to choose something and commit
to it.
I wish it was generally more receptive to that idea of significantly switching gears when it seems like the right approach, but it’s relatable how they react defensively to “maybe we should switch gears and abandon your work thus far”. I know I would personally push back in that situation.
The tendency of coding agents to be attached to their own work is pretty human-like to me. Like no developer wants to be told or arrive at the conclusion that their work on some task was for naught and we need to go an entirely different approach and need to scrap the work.
GitHub lets you comment on code unrelated to a PR’s changes now, instead of it needing to be within ~3 lines of some change. Such a small thing, but a big improvement of code review UX 💬
I think having a plugin system in the app makes more sense than “fork and go”. Then you could extend it for your own use cases, without going through the logistics of git management related to the main product’s development.
Probably my favorite cartoon scene of all time
It's not much different than using products like Loveable. The difference here is the iteration is being done where the code lives (in this case in GitHub), so it's in the perfect spot to be reviewed. And the process is much more customizable in this case using a programmatic system of prompting.
Set up a multi-role Claude workflow system in CI integrated with the batteries-included Claude GH action, then each Discord/Slack/Teams user becomes a leader of a team of coding agents working together.
Claude Code's integration with GitHub Issues and PRs is very impressive on its own. To connect the dots you just need some glue to make the important iteration bits to be a nice experience on the chat side, e.g. screenshots taken with playwright/webdriverio mcp in CI, and preview deployment links.
This is amazing! Very innovative idea. Are you able to share the repo? I'm working on some partyserver stuff and would love to see what you've gotten working for your app 🚀 Thanks!
Shouldn't everything be served over TLS by default by now? Seems like a design flaw that this was not made the case
Anyone using pnpm's "catalog" feature? It's been a game changer for keeping common dependency versions DRY in a monorepo #nodejs #webdev #javascript #npm #pnpm
Shoutout to the devs maintaining @kysely.dev 🧑💻 🚀
There's a fairly novel accidental feature of the library, and while this wasn't intended to be supported (which eventually caused it to be dropped), the devs listened to the community and developed an elegant way to continue supporting it. Thank you!
AI agent “rules” seem very powerful, but I think we need a common file name and format for the different editors/agents. Hopefully the editor creators agree and conform to something, so projects can specify the rules once and they work for all editors.
