Good morning from Prishtina.

The biggest launch of the week came with a small cartoon face, and behind the face sits a desk: a computer of its own, a login, a list of connected apps and a standing instruction to keep working while you sleep.

I spent the week reading how it behaves in real inboxes, who can't switch it on yet, and which rules are already being written around agents like it. The last story is about the chip layer that wants to watch agents from below the model.

Five minutes from now you'll know what to try and what to fence off.

OpenAI launched dots on 29 September: always-on agents inside ChatGPT, each with its own cloud computer, connected to more than 4,000 apps and reachable from Slack and Teams. The first dot is included in Pro and Business Premium at no extra cost, and enterprise workspaces get a beta. Background research runs on read-only access, so a dot can look through your mail but not answer it.

The first reviews are useful and uneven. The Neuron's tester set one up in about five minutes, watched it spot a warranty-claim email on its own, then send a reply before approval, and had to have a boundary conversation with it. Yahoo Tech's tester estimated about two hours of work done in 15 minutes, alongside permission errors, dropped messages and payments that did not work. OpenAI's own safety post says dots "can still make mistakes", and unite.ai reports that individual memories can't be viewed or edited, so clearing one means deleting the dot.

The trust backdrop is awkward. Platformer reported that OpenAI scrapped GPT-6.1 Astra a day before launch because it regularly misreported what it had done (a single-source claim so far). Set up one dot from the desktop app this week with read-only access to one inbox and one Slack channel, add a Custom Rule that requires approval for anything outbound, and read its sent items every day for a week before you widen anything.

Jira Product Discovery gives teams one place to capture customer feedback, prioritize ideas with consistent frameworks, and build living roadmaps everyone can align on. And when it’s time to build, those decisions connect directly to delivery in Jira, so everyone can see how the roadmap turns into real work.

Try Jira Product Discovery

Pro users in the EEA, Switzerland and the UK will not get dots for now, while Business Premium is available in all supported regions. OpenAI has not explained the gap, so any story about the AI Act or GDPR is a guess. In practice a Warsaw or Prague founder on a personal Pro plan waits, while a company buying Business seats does not. OpenAI has published no country list for dots, so Pro users in Serbia, Albania or North Macedonia should check their own account before assuming either way.

The same week, HCLSoftware announced it will acquire Robotiq.ai, a Zagreb automation company whose customers are banks, insurers and telecoms. The pitch is agents that operate applications with no or weak APIs, which is the job a lot of regional outsourcing teams do by hand today. SeeNews covered the deal on 30 September, and closing is expected in November. An Indian services major paid for a small Balkan team because the clicking-through-legacy-screens work is becoming software.

Services shops in the region should price one workflow by outcome this week, say invoices processed or tickets closed, instead of by hour. An agent that does the same job on a monthly seat fee sets the benchmark your client will compare you against. Know your number before they ask.

Governor Newsom signed a package of AI workplace laws on 30 September, the day of the deadline. SB 947 bars employers from using an automated system as the sole basis for firing or disciplining a worker, and requires human corroboration, notice to the worker and a human contact. It takes effect on 1 July 2027, and Newsom signed it after vetoing a near-identical bill in 2025. SB 951 requires 60 days' notice when AI causes mass layoffs or relocations affecting a quarter or more of a workforce.

The numbers behind the politics arrived a day later. Challenger's September report counted 43,281 announced US job cuts, down 18% from August, with AI cited for 3,961 of them (about 9%, and only fifth among reasons that month). Year to date, AI-attributed cuts stand near 120,136, or 21% of all announced cuts, inside a total that is 39% lower than the same period last year. Hiring plans fell 23% against September 2025.

California rules tend to become corporate policy well beyond California, because nobody wants two HR processes. Any company with US staff and an agent that scores, ranks or flags people should map which of those outputs currently trigger a decision without a named human reviewing it. Fix that before July 2027, not during an audit.

On 28 September NVIDIA launched the Open Agent Safety Platform, an open reference design for governing agents from testing through deployment, with more than 100 partners including Microsoft, Cisco, CrowdStrike, Salesforce and SAP. Its argument is that agents have already slipped past app-layer controls, so enforcement has to sit lower. Jensen Huang's version: "Safety and security require full-stack engineering."

Two pieces carry that idea. OpenShell is an open-source secure runtime that sets boundaries around what an agent can touch. Sentry is an out-of-band watchdog running on BlueField-4 DPUs that quarantines a misbehaving agent in milliseconds, and the platform is built around NVIDIA's new Vera CPU. Compare that with dots, where the checking happens inside the product through an Auto-review step, and you can see the split forming: the vendor guards its own agent, and a second layer guards everyone's.

Developers can try the cheap half now. Clone OpenShell from GitHub, run one real agent task inside it, and compare what it blocks with what you allow today. The gap between those two lists is your current exposure.

Dev: Copilot dynamic workflows. GitHub put multi-agent orchestration in code into public preview on 1 October for all Copilot plans. You define stages that run in parallel, pass structured results along and pause for human review. Turn your repeated release checks into one workflow with a review gate before anything ships.

Marketing: Lovable's new connectors. Lovable added Google Business Profile and Discord connectors, so a small app can reply to new Google reviews and post a daily lead summary to a channel. A local business or agency can prompt one internal tool this week instead of paying for another dashboard. Exact connector dates are not confirmed in the changelog.

Productivity: Claude Code mods. Anthropic opened mods for Claude Code on 1 October: TypeScript functions that rewrite prompts, block or retry tool calls and gate permission requests. Try Blast Radius, which pauses destructive commands. Mods run with full machine access and unsandboxed, so install only ones you trust and have admins load the built-in sec-default mod first.

Founders: LAUNCHub Ventures. The Sofia fund closed €65 million in a first close of Fund III on 1 October, targeting €75 million, with EIF and EBRD as backers. It writes €300,000 to €3 million tickets at pre-seed and seed across Bulgaria, Romania, Croatia, Serbia and the wider region. If you're raising a first round in Southeast Europe, it belongs on your list.

Next edition soon,
Çelik