Dagens Vibes — 11. oktober 2026

Fra X-feedet

Software skal læse sine egne fejl

Boris Tane vil lukke løkken mellem produktionssignaler og coding agents: logs, brugeradfærd og support skal tilbage til den, der skriver koden. Menneskets vigtigste job bliver at definere, hvad bedre betyder.

boris@boristane

I spent the majority of my career building observability tooling this tooling should be used to close the loop between production and coding agents, in service of self-improving software I put my thoughts down https://boristane.com/blog/self-improving-software-is-inevitable/

♥ 225 · ↻ 30 · 💬 25 · 🔖 211

https://x.com/boristane/status/2108992293042852217

Delt i feedet via @_overment.

En koordinator — flere maskiner

Riley Brown beskriver en personlig Claude-koordinator i skyen med underagenter på sine Macs og forskellige modeller til forskellige opgaver. Det interessante snit er opgaver, ikke fjernstyrede tastetryk.

Riley Brown@rileybrown

If you want to feel what it’s like to have a personal agent platform created by @AnthropicAI go to the Claude app right now and create a new Claude Code Project, name it Tony or something. Use this prompt. “this project is responsible for moving my life forward, you are to help me get things done in order of how important they are. Analyze all my connections including email, slack, and any platform and suggest threads and automations that we can set up now. One of your threads should be called the Organizer, which does the prioritizing and planning, which keeps an up to date editable document of our todo lists and important documents.” Projects is very good at delegating. That’s what it was made for. There’s a coordinator chat, and then threads (sub agents). What’s cool, is these sub agents can be cloud or local. And since I have my MacBook and 2 Mac minis running 24/7 some of the threads run there. But the coordinator thread is in the cloud so it’s always available. Your coordinator thread can be Opus 5.5, but threads can run different models. Just ask for haiku or sonnet if they are responsible for very routine tasks.

♥ 1220 · ↻ 49 · 💬 97 · 🔖 2366

https://x.com/rileybrown/status/2108899765320638935

Proaktivitet, der faktisk fortjener en push

Michelle gav flere agenter samme mail og kalender: dot fangede en forkert dato, Muse fandt støj. At sammenholde to kilder slår at annoncere, at man har læst én.

michelle 👻@hazelcough

Which of my personal agents was proactive in the right ways? Most of my friends wouldn’t trust an AI agent with access to their personal email and calendar. But in service of laziness (and science!), I gave a few agents read access to mine to see what they'd do without prompting. ✅ dot caught that I'd put a doctor's appointment on my calendar for the wrong day. It noticed that an appointment confirmation email I’d just received did not match what was in my calendar. ❌ That same day, Muse sent me a redundant notification about a stranger tagging me on someone else's GitHub gist. My GitHub handle is @.michelle, so I get mistakenly tagged by strangers all the time. Muse also surfaced the appointment confirmation email, but missed the date mismatch. 🟡 GrokBot hasn't notified me about anything to date. To be fair, I haven't asked it to, and silence is arguably better than what Muse is doing. Useful proactivity > leaving me alone >>> noisy proactivity. It was cool that dot looked across multiple sources to figure out what was worthy of a push notification.

♥ 97 · ↻ 7 · 💬 19 · 🔖 52

https://x.com/hazelcough/status/2108713312729153536

Delt i feedet via @jeff_weinstein.

Pi Durable går i kundeservice

Poppy-demoen lader en personlig agent tale med en butiksagent og simulere en ordreombytning med brugerens godkendelse. En konkret anvendelse oven på Pi Durable — ikke endnu et protokoldiagram.

Brendan Irvine-Broque@irvinebroque

built myself a little demo of the new Personal Agent Protocol (Poppy) using Pi Durable, deployed to Cloudflare ofc ask it to exchange order #1042 https://poppy-caller.roundtrip.workers.dev/ https://github.com/irvinebroque/poppy-demo

♥ 214 · ↻ 23 · 💬 19 · 🔖 265

https://x.com/irvinebroque/status/2108673648999723482

Delt i feedet via @jeff_weinstein.

Haiku 5.5: et første check mod Luna

Morgan Lintons foreløbige Frontier v4-kørsler giver Haiku 56 % ved lavere effort, mens Luna fejler alle opgaver ved low og medium. Luna vinder på pris; Haiku leverer det stærkere første vibe check efter lanceringen.

Morgan@morganlinton

Okay, thought I'd share some early results from my Haiku 5.5 benchmark with @VulcanBench And I figured, comparing to GPT-6 Luna was probably the best comparison. Max is running now, but so far Haiku is doing much better than Luna on accuracy. Luna wins on cost but Luna actually failed every single task in Frontier v4 at Low and Medium efforr vs. Haiku which got 56% at lower effort levels. At Extra High Haiku almost aces Frontier v4 vs. Luna which got less than 50% at the same effort level. More to come, live long and benchmark 🖖

♥ 66 · ↻ 2 · 💬 14 · 🔖 16

https://x.com/morganlinton/status/2108982574693650652

En længere cache er ikke automatisk billigere

Dan McAteer målte sit eget subagentmønster og gik tilbage fra én time til fem minutter. Hans script regner på pauser og dyrere cachewrites; Claude Code-indstillingen for én time hedder "1h".

Dan McAteer@daniel_mac8

Claude Coders: run this script to estimate if switching to a 1h subagent cache lifespan will save you $$$. In my case, the verdict is no. I switched it back to 5m. Script below. Run it to see your results.

♥ 158 · ↻ 10 · 💬 22 · 🔖 221

https://x.com/daniel_mac8/status/2108971441102172620

Papir foran skærmen, AI ved siden af

Noah Tye regner matematik på papir og sender dokumentkameraets billede til Astra med et lille vibecodet værktøj. En god AI-grænseflade kan også være et stykke papir.

Noah Tye@noahlt

Lately I’ve been learning math with ChatGPT (Astra) as my tutor, using this document scanning camera and a vibecoded tool that pastes the captured image right into the chat. This way, I just work problems on paper and ChatGPT gives feedback. It’s really nice.

♥ 9072 · ↻ 520 · 💬 219 · 🔖 5338

https://x.com/noahlt/status/2108675665373589841

Delt i feedet via @lindsmccallum.

Nyhedsbonus

Privatliv er blevet agenternes salgsargument

The Verge gennemgår OpenAI og Metas privatlivsløfter: Muse er isoleret fra andre brugere, men Meta kan stadig tilgå data, og modeltræning er slået til som standard. En privat VM og privatliv er ikke det samme produkt.