Started with one annoying Python typing problem.
Go-brain had me assuming the interface needed a breaking change.
It didn’t.
The agent taught me a better pattern:
TypedDict + @overload + Literal[...]
Same runtime function, better static help.
https://t.co/iiuqUd45EI
just cut a v0.6.0 release for moat:
wiki links, date/draft frontmatter, RSS feeds, page listings, and date-aware nav sorting. Better building blocks for docs sites, changelogs, and knowledge bases.
https://t.co/qN5YLG81IH
A nice harness engineering example:
Start with a small intent: “can I get a feed of stars across my repos?”
End up with reusable software:
GitHub Action
CLI/library
live site integrations
Discord notifications
docs + cleanup
Wrote about the session here:
https://t.co/NleRXqhaCZ
Traders and investors have their own preferences on what they want on their screens. An intraday scalper will value speed and a quick order window above all else, while an options trader may always want the option chain open.
Kite's classic web UI was designed to work universally for everybody. More than a decade later, despite us adding a ton of features and capabilities, Kite is still extremely light and fast with a tiny bundle size.
While desktop trading platforms have historically had widgets, web technologies weren't suited for this for the longest time. We've prototyped this on and off over the years, and last year building a web-based trading terminal became doable with native web technologies.
Introducing Kite Terminal mode. We rebuilt Kite web completely and made every single feature it has into a widget. You can now drag, drop, and rearrange all the components inside Kite to design your own interface tailor-made for the way you trade and invest. We have put in a ton of work to make the interface snappy with smooth UX. Now you can build your own trading interface like assembling Lego blocks.
In addition to investment and trading widgets, we've also added a bunch of productivity and personalization ones like notes, clocks, calendars, wallpapers, my Twitter feed (follow me on Kite) 😛 and more.
I'm excited to see how people design their layouts. Share a screenshot of your workspace in the comments.
and if you need an OSS variant of the same thing - i've built hodor. Uses the excellent @badlogicgames 's pi-sdk, supports gitlab/github, and you can use any model (even bedrock with custom profiles to track costs)!
You can also load custom "skills" per repo/CI job - enforce your own conventions, security rules, or review checklists.
This is what we @zerodha have been using for the past few weeks and it has caught some really non-obvious subtle edge cases/bugs.
https://t.co/EInC9k7JL7
@tmr11235@badlogicgames@nicopreme Interesting, I've tried the /tree but what kind of instructions in my agent would be useful to get the agent to load the skill automatically, wasn't able to get it to load the skill to see it in action
@shantanugoel@thekitze Also WIP, but found that taking a much simpler approach with a basic rule engine on when to run with a heart beat to check seems easier for the agent to extend
https://t.co/2uZLkV88d3
https://t.co/g2uw8lHkL2
2023: LLMs generate random strings of text and "hallucinate."
2026: Linus Torvalds, creator of Linux, says LLMs can generate code better than he can write.
It's remarkable how good these large language models have become in such a short span of time, to the point where they're on par with, or better than, humans across numerous domains.
K (Kailash) recently wrote a brilliant post about how software development, as we've known it for decades, is over. He's far from alone; many people at the forefront of technology are saying the same thing. And replace "software" with almost anything else, and the statement still holds. What are the new frameworks and methods going to look like? 😬 Here's an excerpt on hbeen ow LLMs are helping him:
"As a developer with a bottomless wishlist of things I wished I could have done or tried, I've been able to use LLM tools to not just rapidly prototype and validate complex ideas, but actually write good quality, production-grade software (my own subjective metric, of course) with better code than I could have written manually. Things where I knew exactly what had to be done but was constrained by physical limits, and also things that were unclear to me and needed novel ideas, approaches, and leaps. All the while, learning and deepening my own understanding.
The physiological, cognitive, and emotional cost I generally incur to achieve the software outcomes I want has undoubtedly reduced by several orders of magnitude. The time and bandwidth this has freed up, I now spend on engineering, architecting, debating, tinkering, expanding my imagination, and writing much more concise and meaningful code, the code I actually want to write.
Remember the old adage, "programming is 90% thinking and 10% typing"? It is now, for real."
What's made this real for me is watching Karthik build his own website, complete with a quiz, with zero programming knowledge. This is a guy who used to play words like "Dog" and "Fan" in Scrabble. I now have full FOMO. 😄
Links in the comments.
@ghoseb@badlogicgames@mitsuhiko@nicopreme Nice, that's interesting, although i haven't had such a usecase yet for this but i can imagine that this would be useful. It's wild to observe how agents are able to coordinate so well
@zats@badlogicgames@mitsuhiko@nicopreme I spawn new instances manually in tmux and they auto register to the mesh or agents spawn dedicated agents using a spawn_tmux_session tool to consult our delegate. I see that these agents self organize as they are aware of mesh participants
Three months ago we all built durable execution engines, now we're all building sandboxes. Let's make some predictions what we will build three months from now!
My recent favorite prompting technique:
<give the agent a list of bugs>
Then:
```
for (bug) {
do triage
explain bug and issue based on triage
interview me what to do
update plan
}
```
You must add an instruction for "interview" to use your agents question tool.
Built https://t.co/9B0BKb4G3u yesterday in a single day - a jq/yq-like query processor for HUML. had the LLM study yq's test suite, generate tests, then implement against them while i mostly watched. wrote about the workflow here.
https://t.co/Uws0yvxFw7
I wrote a similar alias using @badlogicgames pi-mono:
```
howdoi() { pi --no-tools --system-prompt "Output a shell oneliner. No markdown. No hardcoding." -p "how do i $*"; }
git status | howdoi cleanup the vim backup files
find . -name '*~' -type f -delete
```