There's a difference between AI sidekicks and background agents. AI sidekicks require you to prompt them to handle tasks on the fly, and don't work without prompting. They help you move faster through your work, but don't fully take work off your plate. AI sidekicks bring between 10-20% efficiency for the average desk worker, and can cost a fortune if mismanaged. AI background agents run completely autonomously, only surfacing to you for exceptions, judgements, and other human-in-the-loop events. They actually remove work from your plate, and the workflows we take on see between 60-90% efficiency gains. The enterprise value ratio between the former and the latter is 1:10. Yet there's a ton of sidekicks like Claude Cowork, Codex, and GitHub Copilot, and almost 0 AI background agents. This is because setting up AI background agents is incredibly involved. You have to: - Go deep into a company, and understand how processes run today, and how exceptions are handled - Re-engineer those processes so that they are reimagined for an AI-native future. For example, 15-step workflows can now be condensed into 5-steps, and AI can handle 4 of them. - Build the agents custom for this process, on top of the systems of record. If your work runs on Dynamics 365 and Salesforce, your agents should do the same. - Integrate human-in-the-loop feedback and model-optimization so that the agents get smarter, faster, and cheaper over time. This is nearly impossible to do alone, because access to frontier AI talent is limited to top tech companies. And MBB firms like McKinsey will pitch you on AI transformation just to hand over a 300 page slide deck after 6 months and not ship a single agent. There's only one company on the planet that will do this work for you, and do it well, and that is linked below.
Dagens Vibes — 10. august 2026
Dagens feed peger samme sted fra tre vinkler: værdien flytter fra chatbotten til arbejdet omkring den. Agenter skal ind i rigtige processer, computer use skal kunne lukke hele loopet, og selv lokale modeller bliver først spændende, når økonomien og installationen hænger sammen. Mindre demo, mere drift.
Fra X-feedet
Den tungeste pointe er organisatorisk: sidekicks gør mennesker hurtigere, mens baggrundsagenter kræver, at selve workflowet bygges om. Procenttallene er salgsbrochure, men skellet er præcist — især i sagsbehandling. Hvis tokenforbruget falder til nul, når medarbejderne melder sig syge, har man stadig bare en meget dyr autocomplete.
DHHs Fable-forsøg er et ordentligt datapunkt: en Python-library blev omskrevet til Rust i ét 11-million-token-løb, med 2 ms opstart, 9,6× hurtigere rendering og én dependency-fri binær. Det interessante er ikke “Rust fast”; det er, at en agent kan tage en afgrænset rewrite hele vejen til et målbart bedre artefakt.
Fable one-shotted a Rust rewrite of the TerminalTextEffects Python library in 11M tokens. Startup time went from 87ms to 2ms and rendering speed is up by 9.6x. Now zero dependencies and a 3mb single exec 🤯 https://t.co/3cTEQAqYdO
Computer use bliver konkret, når den ikke stopper ved “jeg fandt fejlen”. Her læste Codex crashlogs fra en macOS-beta, diagnosticerede problemet og udfyldte Feedback Assistant med beskrivelse og relevante logs. En lille use case, men et komplet loop — og derfor mere interessant end endnu en browseragent, der kan bestille sokker.
Codex investigating a macOS beta crash and filing a bug report with Apple. Yet another awesome scenario for computer use!
Just had a bug show up on my MacBook Pro and I needed to reboot, so I had Codex look at the logs, figure out the issue, then submit a bug report in Feedback Assistant with the proper logs and a detailed description There's something super cool about this idk why https://t.co/djT3c7YHAB

FrankenTTS er dagens maker-projekt: Qwen3-TTS porteret rent til Rust og kørt lokalt i browseren via WebAssembly, inklusive voice cloning. Browseren er ærligt langsom og sluger cirka 2 GB model, mens den native CLI er hurtigere end realtid. Ingen GPU, ingen server, ingen “upload lige din stemme til denne tilfældige startup”.
I'm pleased to introduce: https://t.co/onVx6sUBGz It's my new FrankenTTS project delivered entirely in the browser using WebAssembly! It's much slower than real time (unlike on the Mac, where it's now faster than real time), but it actually does work! And I'm in the process of further optimizing the performance and ensuring that it can run on an iPhone. This site goes well beyond other client-side browser TTS demos because, in addition to being able to use the several preset voices, you can also record yourself directly on the page and very easily make a voice profile, then generate clips using it. You can even share the profile you generated along with the text to read out loud, with both stored entirely in a compressed URL string. The voice profile is just a vector of 1,000 numbers, but does an incredible job of characterizing a voice. There's no real web server here; the whole thing is static html/js and served with Cloudflare Pages (the full source for the site is included in the FrankenTTS repo so you can inspect it all). Beyond that, I've been pushing awesome new improvements all day, including a mini transformer-based noise-reduction stage that leads to dramatically better sound quality with your enrolled voice sample, even when using an iPhone voice memo or built-in MacBook microphone. It also has improved console output and automatically keeps the model resident in memory for 10 minutes after you do a generation so that any subsequent generations using the ftts command are much faster. If you don't generate anything for 10 minutes, it automatically unloads it from RAM. This went from a quick-and-dirty, fun project to what I think is basically bleeding edge for something that can run this easily on super common consumer hardware without needing a GPU or complex setup. Give it a try, it's very easy to get started on a Mac (it also works on Linux and Windows, btw), and if you're willing to wait longer for the generation, you can now do it all from the browser!
Den kolde spand vand til hjemme-datacenteret: en gennemsnitlig OpenCode Go-bruger brugte 1,14 dollar om dagen på DeepSeek Flash V4, mens et tilsvarende dual-DGX-setup koster 10.000 dollar. Lokalt giver stadig privatliv og kontrol; “jeg sparer penge” kræver bare omkring 24 års stoisk dedikation.
the average OpenCode Go user spent $1.14 per day on deepseek flash v4 this past week the dual DGX setup people are running to do the same costs $10,000 it takes 24 years to break even at 10x the usage it takes 2.4 years
there are a lot of reasons to want to run models locally cost is probably not one of them
Pi fik en kærlighedserklæring som det mest fleksible og billige agent-stillads: alle modeller, alle providers, open source og bygget til at blive ændret. Ikke ligefrem en uvildig benchmark, men argumentet rammer det, Pi faktisk er god til: primitives frem for et forseglet produkt.
Pi - best harness for customisation - best building block of all the harnesses - lowest cost per task - fastest performance - supports all models and providers - free to use - open source I love https://t.co/R3ndz7df8b https://t.co/Hz9fXEG1Qu

Dagens lille prompt-gem: bed Claude skrive i ASD-STE100, også kaldet Simplified Technical English, og giv den Zinsser-reglerne simplicity, brevity, clarity og humanity. Det er ikke magi, men det er en langt bedre stilkontrakt end “vær kort og professionel”, som mest producerer LinkedIn med slips.
Oh my god. It's hard to express how big a positive impact this tip can have on your life if you talk to Claude a lot. > Tell Claude to write in ASD-STE100, or Simplified Technical English As a bonus, tell Claude to follow Zinsser's four principles of quality writing: 1. Simplicity 2. Brevity 3. Clarity 4. Humanity h/t @DanielLockyer

Nyhedsbonus
Anthropic gør Claude Codes auto mode til standard for Pro, Max og Team fra 14. august. I stedet for konstant godkendelse stopper den kun irreversible, destruktive eller eksterne handlinger. Det opsigtsvækkende tal: auto mode fangede 89 procent af farlige handlinger i testen, mens mennesker fangede 13,6 procent — vi godkender åbenbart permissions som cookie-bannere.
På Def Con demonstrerede noRecognition et genereret mønster, der fik en Flock-kameraopsætning til ikke at detektere en mønsterindpakket bil. Projektet er trænet gennem 31 millioner tests mod 11 detektionsmodeller. Kameraet optager stadig; algoritmen glemmer bare at råbe op. Cyberpunk-modeindustrien har endelig fundet sit egentlige marked.