@GeoffreyHuntley agreed, something like that will probably be main "programming" paradigm in a future. live apps reacting to user demands. I was thinking about doing it in .net, but it's not the best platform for this kind of stuff. lisp looks much better.
@GeoffreyHuntley@Scrubghetti@theo sometimes agents just do stupid shit. it's essential to have another agent review the work, no matter what model is selected.
@GeoffreyHuntley I automatically attach every new codex thread to a specific project. this injects a history summary + latest messages. the context is bound to 8kt and updated in the background. there's additional infrastructure (jobs/notes), but it's basically a single sqlite db and a plugin
@GeoffreyHuntley@temporalio scheduled predefined agents analyze project structure and create tasks. each task is an owner/dev/qa workflow controlled by the owner. schedule provides a pseudo-loop, and tasks are flexible workflows. don't make workflows rigid, it will inevitably fail.
@GeoffreyHuntley It will continue making sense until training set stops being relevant. Agents are much more effective with TS/React because there's so much training data.
I imagine, after a while, there won't be any code, just **live** apps with runtime coding hidden from humans.
@GeoffreyHuntley funny coincidence, I'm planning to release a runtime patching dotnet lib soon. 100% written by my own workflow-based AI agent, win/linux/macos. just need some free time to put it all together as a product. thinking of targeting it for testing at first.
retweets appreciated
hi folks, some important life news
i’m looking for a new employer - contact via [email protected]
this is your chance to nab someone
who is nine months in with knowledge
on building agents, is a teacher and
prolific author and public speaker.
I’m based in Sydney, Australia but work remotely. Role guidance IC5. SF.
@pierceboggan@itsmattchan@BrigitMurtaugh Just a small example so I don't look like a whiner.
This suggestion keeps appearing, no matter how many times I go back to the previous line and then press Enter. So there's enough time for a "proper" suggestion. And this suggestion quality is somehow the norm lately.
@pierceboggan@itsmattchan@BrigitMurtaugh Honestly, tab completion is more of a nuisance that I have to fight now. It was great a year ago. How did better models result in worse tab performance? Honestly, as a paid client since beta, I'm quite mad about this.
@GeoffreyHuntley@AmpCode@_dhamidi I think I already have an account at [email protected]. Just didn't get a chance to properly try it out. But credits would be nice, thanks!
Unfortunately I don't know @_dhamidi. I'm a lone consultant, don't really know anybody 😀
@GeoffreyHuntley@AmpCode Modes are nice because they are like hardcoded prompts with a designated name. You can reference that name in other modes/prompts, and be sure that LLM won't do anything stupid (since all available modes are listed in a system prompt)
@ErikEJ@jsuarezruiz I tried it this weekend with .NET 10 Preview 4. Had an issue where .cs files in the same directory didn't get picked up. I needed to share a few util methods between multiple .cs scripts.
What's going on with GitHub Copilot in Rider? Lately, it has become completely unusable. Suggestions are missing or plain incorrect.
@JetBrainsRider@github