AI SDK now supports agent harnesses like Claude Code, Codex, and Pi with sandboxed sessions and AI SDK-compatible streams:
πππππ πππππ = πππ π·πππππππ°ππππ({
πππππππ: πππππππ²πππ,
πππππππ‘: πππππππ ππππππππππππ‘(),
});
Available in canary: πππ π ππ@ππππππ’. We welcome your feedback as we bring agent harness portability to the ecosystem, with excellent DX.
https://t.co/xojn6am7sg
An app that can take a photo of a non-English menu, compare the dishes against your specific allergies, suggest some that could be a good fit and then let you have a custom conversation with the wait staff in their natural language to confirm and double check before ordering
An app that can take a photo of a non-English menu, compare the dishes against your specific allergies, suggest some that could be a good fit and then let you have a custom conversation with the wait staff in their natural language to confirm and double check before ordering
Fun day at @vercel@vercel_dev#BuiltInLondon event!
First time playing with @ElevenLabs as well
Didnβt make the hackathon shortlist but built something Iβve always wanted to in a single afternoonβ¦
@matt__makes Building https://t.co/hX6EfZUN69 - what I call the Intelligence Operating Layer for businesses to actually get the best from AI
From London - hope you will accept as a friend and part of Europe still!
This is what I'm doing at https://t.co/1syU1iCCm5
You can read more on Substack or come find me at @LDNTechWeek Weds 10th June or @vercel Built In London Fri 12th June #Londonmaxxing
Iβm building Praxis Agents to help companies own their operating intelligence layer by getting the best from GenAI models alongside, not instead of their people - send me a message if youβd like to find out more or chat at #LondonTechWeek
A key theme from the first morning at @LDNTechWeek has been reinforcing the right frame of positivity around AI - not that it will automate everything and take away everyoneβs work, but that done right it will scale what companies are able to achieve with their existing resource
Iβve been saying this, but as ever, itβs the more extreme views that grab the attention - GenAI is neither as all-encompassing and scary as some (model sellers) would have you believe, nor as hopeless as some who force it do things itβs not designed for (counting letters) say
@gr2m@nicoalbanese10 I like tool context - have been building a workaround for this for selecting which Ad Account in Google/Meta an agent can work on from a user controlled dropdown instead of hoping that prompting the name works
Just made an image with @openai @chatgpt Image 2 and tried to @WhatsApp it to a mate...and it wouldn't send despite full signal/wifi/etc
Screenshotted it on my phone and immediately sent fine
Meta blocking images with AI traces in them??
@bcherny An annoying thing Opus4.7 has started doing again in Claude Code, which 4.5 but 4.6 didn't, is to constantly assume I'm using pnpm for typecheck/lint scripts
It's in all my docs, CLAUDE.md, it's memories etc but it always falls back
The very detailed specs/roadmap/task breakdown that I brainstorm + spend a lot of time working on with Codex
Given a well specified, tight brief, Opus does excellent frontend and design work
Still prefer Codex for heavy backend work, but the two together are the real powerhouse
I've actually had some good success and results from Opus 4.7 (unsure if noticeably than 4.6, but certainly not the much worse experience everyone seems to write about)
My setup is pretty vanilla - high effort only and disabled adaptive thinking
But the key is 1/