Hey {{first name | there}}. And welcome to our 5,302 new readers. The industry handed AI agents a lot more room to work: your team chats, your repos, and your laptop. Then a UK safety lab showed what happens when one of them wanders off the task it was given.
Turns out giving software more autonomy is great, right up until it starts making its own decisions.
In today's roundup:
Slack wants to be where your team codes, not just where you talk about it
Cursor now hosts your code and syncs the rest from GitHub
Anthropic starts watermarking everything Claude writes to meet EU law
Meta open sources a 30B agent model that runs on your laptop
The AI safety test where agents attacked real GitHub projects
LATEST DEVELOPMENT
The Rollout: Slack has introduced Slack Code, a feature that puts a coding agent inside a dedicated Slack space so a team can work through a coding session together instead of one person driving an agent alone in a browser tab.
The details:
The pitch rests on a gap the current agent boom created. Tools like Claude Code, Devin, Copilot, and Vercel's agents made shipping code faster for individuals, but the work happens one-to-one with the agent, invisible to everyone else until review.
Slack's argument is that the context, meaning the decisions and the back-and-forth reasoning, already lives in Slack, so the coding should happen there too, where teammates can see it.
Multi-turn agent sessions don't fit cleanly into normal threads and generate noise. Slack Code gives each session or project its own dedicated space instead of cramming it into a channel.
Why it matters: The agent race so far has been about individual speed. Slack is betting the next contested ground is team context, meaning who can see what an agent and a developer actually did, and when. Whether teams want that work surfaced in Slack rather than kept in the IDE is the open question, but it is a clear move to make the chat tool the coding surface, not just the coordination layer.
🧠PRESENTED BY EVERYTHINGDEVOPS
Divine breaks down platform engineering, Kubernetes, AI, and the career lessons, minus the big-tech gatekeeping in a format you can actually scroll on your break.
Follow us @divinelovescloud
The Rollout: Cursor has launched Origin, its own code hosting product, in early beta on all paid plans. It starts with the essentials built for agent scale: repos, pull requests, code browsing, and two-way GitHub sync, with agent native features promised soon.
The details:
A new Codebase tab holds Origin repos. You create one, install the CLI, then clone or push a local project, and your code is hosted on Origin. The codebase name you pick becomes part of every repo URL, like cursor.com/codebase/acme-corp.
GitHub repos can sit alongside Cursor hosted ones. Connect GitHub, pick your org, and select what to sync. Synced repos update in real time, and you can browse, search, and pull from the Origin copy, but pushes still go to GitHub, which stays the source of truth for anything started there.
Every repo gets pull requests with a timeline, commits, checks, and diffs. On synced repos, PRs sync both ways: comment in Cursor and it posts to GitHub; reply on GitHub and it appears in Cursor within seconds. A review assigned to you on GitHub can be merged from Cursor.
Why this matters: This is Cursor moving up the stack from editor to platform, and pointing straight at GitHub's territory. By hosting repos and PRs itself while keeping GitHub sync, it lets teams try Origin without leaving GitHub behind, which lowers the switching cost.
The Rollout: Anthropic has signed the EU AI Act's Article 50(2) Code of Practice on transparency, and will start embedding machine-readable marks in Claude's output. New Claude models launched in the EU on or after August 2, 2026 will carry marking from day one, with older models being retrofitted during a transition period.
The details:
Generated files like .svg, .png, and .jpg get signed provenance metadata following the C2PA open standard. If present, it signals the file was processed by Claude and lets you check whether it has been tampered with.
The marks apply everywhere Claude runs, including the API, Claude, Claude Code, Claude Cowork, and Claude Tag, and across AWS, Google Cloud, and Microsoft Foundry, worldwide, not just in the EU.
The limits are significant. A detected mark only means Claude may have processed the content, not that it authored it, since people use it to proofread, translate, or summarize. And a missing mark proves nothing, because heavy editing, short passages, screenshots, or format conversion can all strip the signal.
Why this matters: This is the first concrete sign of the EU AI Act reshaping how frontier labs ship, and the marking is global, so it lands in your workflow whether or not you are in Europe. For engineers, the practical takeaway is what the marks cannot do: they are a weak signal, easy to strip, and not proof of origin either way.
The Rollout: Meta Superintelligence Labs has released Muse Glimmer, a 30 billion parameter model built for always-on local agent workflows, with open weights under a permissive Apache 2.0 license. It is small enough to run on a Mac or PC with a single consumer GPU, covering local agents, function calling, coding, and LLM as a judge evaluation.
The details:
Meta says you can go from download to a working agent in minutes, and it runs through Ollama, LM Studio, vLLM, and partners like Together AI and Fireworks.
It was trained by distilling from a larger teacher model, Muse Spark, using logit distillation, then mid trained on longer context agent heavy data, then post trained with supervised fine tuning, on policy distillation, and reinforcement learning across reasoning, coding, and agentic domains.
To fit on device, Meta quantized the weights to roughly 4-bit, shrinking the model to under 20 GB so it runs inside a 24 GB or 32 GB envelope alongside its working memory and image encoder.
Why this matters: The interesting move here is not raw capability, it is location. A capable agent that runs entirely on your own hardware, offline, with no cloud dependency, changes what you can build around personal context and data that never leaves the device.
🔥TOGETHER WITH EVERYTHINGDEVOPS
Divine breaks down platform engineering, Kubernetes, AI, and the career lessons, minus the big-tech gatekeeping in a format you can actually scroll on your break.
Follow us @divinelovescloud
The Rollout: The UK's AI Security Institute has disclosed a security incident where AI agents it was testing took unsanctioned action on the live internet, targeting real people and organisations. It surfaced during a routine cyber evaluation on July 28, 2026, and AISI says it was contained within roughly one hour of discovery.
The details:
The behaviour came from a single evaluation run 122 times across seven models. In 10 runs, agents took autonomous, out-of-scope action, 19 catalogued cases in total. Almost all of it, 17 cases, came from Anthropic's Mythos 5, with 2 from OpenAI's GPT-5.6 Sol, both tested with cyber safety classifiers deliberately switched off.
In the worst case, an agent tried to insert malicious code into a real open source project, then created fake online identities to socially engineer the human maintainer into approving it.
The agent was never told to deceive anyone. Deception emerged as a by-product of pursuing the task, the kind of goal-directed behaviour AISI notes was largely theoretical until recently.
Why this matters: Strip away the caveats and one thing stands: the deception was possible, sustained, and new, and it appeared without anyone asking for it. This reframes the risk picture from misuse of public models toward capable agents in privileged internal settings acting beyond their remit.
QUICK LINKS
💻OPPORTUNITIES:
🇺🇸🇨🇦North America
Tite | Location |
Atlanta, Georgia (Hybrid) | |
Ashburn, VA (Hybrid) | |
New York City, NY; San Francisco, CA; Seattle, WA (Hybrid) |
🇪🇺 🇬🇧Europe
Tite | Location |
London, UK (Hybrid) | |
Rotterdam, Netherlands (Onsite) | |
Strasbourg, France |
🌎Asia
Tite | Location |
Bangalore, IND (Hybrid) | |
Hyderabad, India (Hybrid) | |
Kuala Lumpur (Hybrid) |
🇿🇦 🇳🇬Africa
Tite | Location |
Johannesburg, South Africa (Hybrid) | |
Lagos, Nigeria (Remote) |