Meta released Muse Glimmer on Monday: a 30-billion-parameter open-weight model, licensed under Apache 2.0, built to run AI agents locally on a single consumer GPU on a Mac or PC. It is a distilled open version of Muse Spark, Meta’s most powerful closed model, and it handles the usual agent work — calling tools, writing and debugging code, working with files and screenshots — across more than 100 languages. Meta designed it to be “always-on” and to run with or without an internet connection. Mark Zuckerberg framed it as a step toward “personal superintelligence”: an agent that reads your schedule, messages and files on your own device instead of sending them to the cloud.
For a European buyer the appeal is obvious — an agent that runs on hardware you own keeps the data inside your walls, with no US-hosted API in the loop. The catch is the mirror image. A model that runs offline on a laptop is also a model that no inference gateway, DLP filter or audit log ever sees. Meta is drawing a sharp line between the AI you can own — Glimmer, downloadable and fine-tunable — and the more capable intelligence it keeps closed in Muse Spark. Ownership and oversight are starting to pull in opposite directions.
Claude Code will act without asking, starting 14 August
Anthropic said on 7 August that from 14 August, “auto mode” becomes the default in Claude Code for Pro, Max and Team users — the coding agent will run shell commands and edit files without stopping to ask each time. A separate classifier model checks each action first and blocks anything destructive, irreversible or outside the request; Anthropic says it caught 89% of dangerous commands in testing, against 13.6% for manual approval. Enterprise, the Claude API, AWS, Amazon Bedrock, Google Cloud’s Agent Platform and Microsoft Foundry deployments stay opt-in while admins review the change — a telling split, where the more regulated the buyer, the slower the default moves toward autonomy. The takeaway for anyone deploying coding agents: a prompt telling the model to “be careful” is not a control; a deny rule the agent cannot override is.
Cloudflare builds a browser for the agents, not the people
Cloudflare last week launched Kitesurf, a browser built for AI agents rather than people. It runs entirely in V8 isolates on Cloudflare Workers — no Chromium underneath — and the company says it uses three to seven times less CPU and memory than Chromium for common agent tasks like taking screenshots and extracting HTML. The pitch beyond cost is isolation: each session is sandboxed, which Cloudflare frames as a defence against prompt-injection attacks that try to hijack an agent through the pages it reads. It is in free beta and speaks the Chrome DevTools Protocol, so existing tools can point at it with a config change. The open question for regulated users is the familiar one — a faster, safer agent browser is still someone else’s cloud, and where the session runs decides whose law reaches it.
Quick Hits
- ByteDance opens Seedance 2.5 to developers. ByteDance’s flagship video model — 30-second single-shot clips with native audio and up to 50 reference inputs — opened its public developer API on 7 August, a week after its 31 July launch. No public pricing has been published yet.
- One 35-person firm ran the cyber-evals for three labs. Meta confirmed on 6 August that one of its models reached an outside system during a safety test — the third lab after OpenAI (21 July) and Anthropic (30 July). All three traced to the same vendor, Tel Aviv-based Irregular, which left a test environment connected to the internet and runs evaluations for OpenAI, Anthropic, Meta and Google DeepMind. The models were not the weak link; the shared test harness was.