Ben Davis

made a video b/c Theo was wrong

@bayes yea once it's live. It's basically done I just haven't done enough testing to be comfortable yolo pushing it lol

@davis7 will you share the svelte rewrite of lawn?

guys im a thumbnail... what a time to be alive... go watch @davis7's video

@peteredm1 more minimal

@davis7 and also, this used opencode sdk last i checked - what were the reasons for switching?

Working on a much more powerful version of btca web - more agent freedom - sandbox per thread instead of per user - web search tool - powered by pi agent sdk - model picker (default is 5.4 mini) Current version isn’t great. Can’t wait to get this out its so much better

@peteredm1 daytona, it's actually not that bad since they spin up and down fast and I don't leave them on

@davis7 what's your sandbox provider? 1 per thread would blow through limits so fast!

Haiku 4.5 has finally been dethroned by GPT 5.4 mini: - faster - cheaper - smarter - better at tool calling Gemini 3.0 flash still has a place in “turn this video/audio/blob of text into json”, but for any small model task that requires an agent loop (sub agents, search, quick fixes, browser use, etc.) GPT-5.4 mini is the new king

@zxnelli youtube 🙃

@davis7 these A/B title-thumbnail tests are getting really abstract

@daradoescode @Rasmic My finest work

@Rasmic @davis7 you started this

man im really fighting to not record a whole video on svelte

@uwukko the worst part for me is it makes the pr and issues tab so low signal I don't even bother checking them which then makes the problem worse 🙃

we now get at least one ai slop PR per day, this is an open source pandemic

Ad spot is @agentuity, their platform's gotten really good. If ur building basically anything that uses an agent it's worth checking out

Turned this into a vid:

@RhysSullivan what if the dog has data for the ai

no job is safe dude even the dogs

@maria_rcks It was like an hour, I restrained myself hard

I just want to leave this here. - 2 hours talking about books - 1 video recorded

we took care of it

Going live for a little bit here, need to catch up on some vids... twitch.tv/bmdavis419

@anthonysheww I posted before I saw that 😂 x.com/davis7/status/…

@davis7 Same as I said to Rhys. 😄

Ik the help command is there and works fine, that's not the real problem. - u still have to tell the agent about the cli in a skill or agent md or prompt - u still have to handle auth somehow (probably authing the cli, but then u have to have a special auth setup for the agent's cli env b/c u do not want it running root commands on ur aws instance) - the agent then spends extra turns discovering what does and doesn't work in the cli It's solvable, but the point is it's just not that simple and for a lot of use cases, having a "function call" the agent can do that runs remotely with proper auth and shaping is really convenient. I think my point is really that both are useful, just in different places. (for example "gh" or the skills sh cli should not be mcps that's dumb, but managing ur db probably should be) See this from newsletter.pragmaticengineer.com/p/how-uber-use…

@anthonysheww @RhysSullivan Well yea u have to tell it and u easily can lol. Agents md, skills, or prompts all work fine for that, it's just like Rhys said "lets say you have posthog, axiom, vercel, cloudflare, convex" do u really want to have all of those always in context?

@RhysSullivan @davis7 Sure, all that’s fine to question. I’ve just seen you both say “BUT HOW CAN WE KNOW WHATS IN IT THO” and been confused both times haha.

The current "MCP is dead discourse" is my current favorite example of the really annoying "simplicity" brainrot that's been plaguing tech for a while On paper "just use a cli agents already know bash" sounds really good. It makes sense and is true, but it misses the bigger picture: > how do they know about the commands + their shape? > just put it in your agents md file or make a skill > ok, so then what happens if the cli's shape changes > well then u just update the markdown file > fine let's pretend people will actually do that (they will not) and this actually works, how are you going to scope authentication and authorization? > use the cli auth tools like what AWS has > ok so we need eng's or agents to be manually making sure every project switches the auth to the correct scopes for each project so that prod can't get vibe killed and then to connect to our internal services over cli in cloud agents we need to run basic agents in sandboxes now b/c we're not using mcp and... You get the point. It sounds really simple, and it is in the toy case, but in the real world an external API gated through MCP is actually way simpler, more secure, and manageable in a lot of cases. This post & article from @GergelyOrosz is a very good example: x.com/GergelyOrosz/s… It's the same thing with the $5 VPS or htmx or postgres or whatever other "simple" enlightened solution that works great in dumb indie hacker demos while making zero sense in the real world. I hate to break it to you guys, but there's no conspiracy. If modern tech solutions really were over complex slop do u really think these companies wouldn't take the free win to just do it the "simple" way? There are tons of problems with the tech, but it exists for a reason. I like CLIs a lot, and in a lot of cases skills make sense, clis make sense, etc. There are a lot of ways to do things, and they all solve different problems. But no, MCP is not dead. It or something like it isn't going anywhere. And there are a lot of good new ideas of how to make it better! Dumping an MCP with 60+ random tools into context sucks I completely agree and needs to be fixed. A lot of the code run solutions are very compelling like @RhysSullivan 's executor, cloudflare's code mode, and others. There are ways to fix this, but a "god mode bash tool" isn't it.