nutcracker

🧑‍💻☁️☁️☁️☁️☁️

[nutcracker @miloskrca]

@badlogicgames 100%
If any thing, in our case, people would start being comfortable really quickly and start asking impossible things. One was: “find all low hanging fruit in the app that would make customer experience better and implement it” 🥲 I loved it

[nutcracker @miloskrca]

@AnirudhWith @badlogicgames @daytonaio @vercel_dev @chatsdk sweet, thanks 🙏

[nutcracker @miloskrca]

@bentlegen @modemdev A lot of how people started using it is something to the tune of: “I don’t want to bother an engineer so please tell me how this system works and how many users are using it”. Engineers on the other hand do not use it as much at all and like their cli tools better 😅

[nutcracker @miloskrca]

@bentlegen Nice! Have you been using it in an actual work setting? Are you happy with the approach you took? I’ve built a similar slack bot a couple of months back but decided to focus it on non-technical users mostly and have seen some really fun usage

[nutcracker @miloskrca]

https://t.co/Hfj5NN9Xja

[nutcracker @miloskrca]

@axhoff @badlogicgames Yes. A basic implementation but it worked from the first try. I was testing Codex 5.3 and decided to build this bot (wasn’t thrilled with OpenClaw tbh) and turns out it is pretty awesome

[nutcracker @miloskrca]

@marcklingen We just built our own on top of claude code some months before openclaw came out. It’s really not that hard. Similar setup to openclaw but we only gave it the skills that make sense for our business and only with very controlled access at that.

[nutcracker @miloskrca]

@stevekaliski Very cool! How big would you say are the things engineers would use minions on? In follow up post it would be awesome if you shared how you manage minion’s context

[nutcracker @miloskrca]

@iannuttall I guess running many agents when you have a great plan cut into small enough pieces can work. Especially as models start generating better and better code. The human behind the plan will still be critical imho, no value in building smth nobody needs

[nutcracker @miloskrca]

@iannuttall Why are you so high on Codex? I saw you mention it to be a superior engineer but I am using both Codex and Claude Code and always get better results with Claude Code. I’m genuinely curious

[nutcracker @miloskrca]

I’d add that using Skills which leverage bash and CLIs is the way to go most of the time when you need an LLM to be able to do something. Bash/CLIs are fast, composable and self documenting, everything MCPs and APIs are not

[Daniel San @dani_avila7]

We saw exactly this in my previous startup

RAG + vector DB gives decent results, but agentic search over the repo (glob/grep/read, etc) consistently worked better on real-world codebases.

We even pushed further: RAG + embeddings + AST + tree-sitter. The quality was excellent

But exactly as @bcherny mentions: staleness and privacy, you need continuous re-indexing, and all the code and embeddings must live on your servers.

In practice, fast models + bash-style agentic search ended up outperforming general RAG search, even if it requires more tool calls

This is what we built 👇

[nutcracker @miloskrca]

@garrytan I get LLMs are getting pretty damn good, but I use it every day with Rails and it is not the best language/framework combination. There are too many ways to shoot yourself in the foot even with so much RoR code out there. Also, we should not be counting LOC/day in 2026, cmon now

[nutcracker @miloskrca]

"while Moltbook boasted 1.5 million registered agents, the database revealed only 17,000 human owners behind them - an 88:1 ratio."

Shocker, totally unexpected

https://t.co/wQCcM9rLhP

[nutcracker @miloskrca]

@iannuttall I actually never do that. I want the Agent to use the skill when it makes sense. So when I notice it didn’t do it, I try to improve the skill’s description

[nutcracker @miloskrca]

@hassiebpakzad @rawert @langfuse Thanks for the link! I checked it out and it looks nice but I found only LLM-as-a-Judge evaluators. Any plans to add custom evaluators that users can configure in e.g. Python? Or even just pre-built "contains"/"exact match" evaluators, something more deterministic.