@ShriramKMurthi@psduggirala I think for syntax errors, yeah definitely. Audience matters *so much* for error messages, so there's a lot of opportunity to customize that and get it wrong in ways that makes things so much worse.
@ShriramKMurthi@samth@dimvar “● Confirmed — the spec wants me to write new tests that work under both backends verifying the behavior is sensible, not just flag-and-ignore. I didn't do that. Let me...”
sigh.
anything that I'm doing that is more than a one-shot query/change/update that I'm not confident will be done by the 40% mark of a single context window
I suspect subagents are involved in >80% of my sessions involving any complex work where state is getting created and updated
mostly architecting and coding, but also most of OpenProse, Inc. runs out of a single git repo so a large share of our (still small, but growing) operations runs in claude code sessions, and all of those use subagents to break tasks down and pass context by reference
“1842 - Ada Lovelace writes the first program. She is hampered in her efforts by the minor inconvenience that she doesn't have any actual computers to run her code. Enterprise architects will later relearn her techniques in order to program in UML.”
I am one of the inventors of the UML and I endorse this statement.
When designing DrRacket, we made sure that we regularly tested on underpowered Windows computers, so that our software would run well on generic public school machines of the era. Likewise, we test Pyret on Chromebooks. Make sure you experience what your users experience.
I guess it's not that weird that agents would have a manner of speaking (“smoke test”, “you're right to push back”). It's the kind of thing you notice on teams and joke about.
It's just weird that we're all talking to the same coworker.
At ICML, Jinwoo Kim will present the first approach for enforcing formal structural constraints like regexes, grammars, and schemas on the output of continuous diffusion language models, entirely without retraining. 1/
the concrete thing i want from the control layer: structured speculation. run N parallel exploration branches, evaluate outcomes, commit to the best. current agent frameworks are fundamentally sequential, but the model cost barely changes whether you explore 1 path or 5. this alone would be a big leap from the current read-call-execute loop
I think it's important that we continue taking a moment occasionally to remember that MACHINES CAN WRITE NOVEL CODE AND IT ACTUALLY MOSTLY WORKS lest we forget how amazing this is