Signal #93: Agents are getting ID Checks

Cloudflare let websites spot and block AI agents acting for users, Kosovo put ID and e-signatures on phones, CrowdStrike and Salesforce paid for the parts that can stop an agent, and an EU court adviser set tests for training on publishers' work. The bottleneck is moving from what AI can do to who trusts it to do it.

Your client doesn't care how their site was built. They care that it launches on time, runs on their own domain, and keeps working long after you hand it over.

With Lovable, you describe what your client needs and get a working version you can show them in the first meeting. Then you take it all the way to a live site on your client's domain, built to be found on Google. Lovable's project monitoring feature spots bugs and suggests fixes, so your name stays on work you're proud of.

Start today and take $5 off your first month.

Build for clients, $5 off

_{^{Discount of $5 applies to first month of any paid plan. You need to sign up via this webpage in order to claim the discount. The discount applies to new and returning users without an active paid plan. Terms apply.}}

Friday again, friends.

On Tuesday, Kosovo's digital ministry put a national ID and an e-signature app on people's phones. You can now prove who you are and sign a document without going to an office. The next day, Cloudflare gave website owners a switch to decide whether an AI agent shopping for a customer may come in. Two unrelated launches, one problem: before anything acts on your behalf, somebody has to know who it is and what it's allowed to do.

I think that's where the AI story is heading. Most of this year was about what models can do. The questions that slow things down now are duller. Who is this? Who allowed it? How much may it spend? Who switches it off when it goes wrong? The companies that answer those get paid.

That's how this issue fits together. Cloudflare's update makes agents their own kind of visitor on your site, so you need a policy for them. Kosovo's e-ID gives local founders a trust layer to build on, in the same week Brussels rated the country's digital progress this year as none. CrowdStrike and Salesforce agreed to pay over $3 billion between them, by reported figures, on companies that run, watch and stop agents. And an EU court adviser, in a case a Hungarian publisher brought against Google, spelled out when AI may learn from someone else's work.

Here's my worry. Every checkpoint makes agents safer and slower, and small companies feel that friction first. I'd still take it. Nobody sane hands an agent the company card without these rails.

If you run a shop or a website, read the Cloudflare piece first. If you build in Kosovo, send the second story to your CTO.

On 4 November Cloudflare added a "Personal Assistants" category to its bot analytics. It covers AI agents acting for a specific person: comparing products, reading support pages, completing purchases. Site owners can now set one policy for all agents and a separate one for each supported agent: allow, block on pages with ads, or block. In early access, Account Abuse Protection shows which logged-in accounts are using an agent, so a fraud team can act on one account without blocking that agent for everyone.

An agent buying for a customer is neither a bot to block nor a human to trust, and from this week your site has to decide which it is. For years the rule was simple: people in, scrapers out. A personal agent breaks that, because it carries a real customer's intent but none of the attention your ads and upsells were built for. That's why Cloudflare's setup now asks whether your pages make money from ads. The obvious objection is that blocking agents means blocking customers. True, and that's the point: you need a policy now, not a reflex. Cloudflare published no traffic numbers, so nobody outside knows how big this is yet.

Open the Personal Assistants view in Cloudflare this week and see how much agent traffic you already get. If you sell online, decide which pages agents can use, what happens on ad-funded pages, and test what your login does when an agent tries it.

Kosovo's Ministry of Digitalisation and Public Administration launched a national electronic identity and digital wallet platform on 3 November, with an e-signature app called iSign. Citizens and businesses can now identify themselves and sign documents from a phone, without an office visit or printed paper. The EU put €2.3 million into the platform, according to EU Ambassador Dirk Schuebel.

A state-backed digital identity is the most useful public infrastructure a small market can give its founders, and Kosovo just shipped one. Remote onboarding, contracts and HR paperwork have been slow here because proving who you are meant a person in a room. A fintech, an HR platform or a legal-tech startup could now, in principle, verify and sign inside its own product. The catch is adoption. A day later the EU's 2026 report on Kosovo said the country "made no progress" on digital transformation this year and pointed to "a limited number of available online services, fragmented platforms, and cybersecurity risks". A wallet matters only when banks, employers and institutions accept what's in it.

If you build in Kosovo, ask the ministry now whether private services can verify an e-ID or accept an iSign signature, and on what terms. The same report asks Kosovo to align its laws with NIS2, the EU cybersecurity directive, so if you sell security or compliance work, that demand is coming.

CrowdStrike agreed on 3 November to buy Fiddler AI and Dreadnode, for $495 million combined according to the Wall Street Journal. Fiddler tests agents before launch and traces why one broke a rule. Dreadnode attacks agents and the MCP servers they use to find weak spots. CrowdStrike also added an AI Kill Switch that cuts a misbehaving agent's network access and stops it relaunching. The day before, Salesforce agreed to buy Cresta, whose voice agents handle full customer calls, for a reported $2.7 billion.

The premium isn't for smarter agents. It's for agents a company can watch, limit and switch off. Cresta is worth that much because it already works in live contact centres for Marriott and Comcast. Fiddler and Dreadnode sell the record and the stop button. Berlin's Pliant made the same bet from the payments side: it raised an $80 million Series C this week, co-led by Speedinvest, to let agents like Claude and ChatGPT use its company cards through MCP, with permissions, spending limits and audit trails. You could say these are US and German deals, far from a 40-person team in Tirana. But testing, monitoring and permissions are heavy, unglamorous engineering, which is what this region's teams do well.

If you build agents for clients, sell the controls with them: a log of every action, a spending limit, a way to stop it. If you buy agents, make the vendor show you how to switch one off before you sign.

Like Company, a Hungarian publisher of ad-funded news sites, sued Google in Budapest over Gemini, and the case went to the EU Court of Justice. On 3 November Advocate General Maciej Szpunar gave his opinion. AI training can fall under the EU's text-and-data-mining exception, he said, but only if three things hold: the developer accessed the material lawfully, without getting round paywalls or bot blocks; the rightholder didn't opt out in a machine-readable way such as robots.txt; and the developer can show it took steps to stop the model reproducing specific works too closely. He also said a chatbot answer can infringe a publisher's rights when it copies an article closely enough to replace it.

The opinion turns "we trained on public data" from a defence into something you have to prove. It isn't binding, no judgment date is set, and Szpunar advised the court to treat the training questions as inadmissible, so the final ruling could skip them. But judges often follow their advocates general, and his three tests read like a checklist. For publishers in our region, robots.txt now looks like a legal tool as much as a technical setting.

If you run a content site, make sure your robots.txt and terms say plainly what AI may use. If you train or fine-tune on scraped European data, start keeping records now: where it came from, how you accessed it, and what you do to stop the model memorising it.

Dev: Revolut cut its AI coding bill about 90% by hosting open models itself. Revolut moved about 2,000 engineers onto self-hosted GLM models on 56 Nvidia B200s, at about $0.30 per million tokens, roughly a tenth of Codex or Claude Code. Vision tasks still go to Gemini or GPT through a router. You won't buy 56 GPUs, but you can copy the shape: a cheap default, a router, and a frontier model only when the task needs it. (The Register)

Dev: Cognition's SWE-2 Flash is free in Devin until 30 November. The new coding model runs about three times faster than SWE-2 at a fifth of the price, in Devin Desktop and the Devin CLI. Put your routine tickets through it this month and note where it fails before you pay for anything. (Cognition)

Ops: Infinum spends €400 to €600 per employee a month on AI tools. Netokracija asked Croatian readers who pays for AI at work: of nearly 400, 63% said the employer covers it fully and 22% pay themselves. Infinum has no per-person cap, and Robotiq.ai spent about 3% of last year's revenue on AI tools and hardware. Use the numbers to set your own budget and a cap before usage sets it for you. (Netokracija)

Funding: Bulgaria's Dronamics is heading to Nasdaq through a SPAC. The cargo-drone maker agreed to merge with GP-Act III at about a $500 million enterprise value and expects about $100 million in net cash at closing, in the first half of 2027. GPAT shareholders vote on 12 November on extending the deadline, which is the first test. Watch how much of the trust survives redemptions before you treat SPACs as an exit route. (The Recursive)

Compute: EuroHPC explains free EU supercomputer access at Web Summit. On 11 November at 16:00 in Lisbon, EuroHPC runs a session on free industry access to EU supercomputers, AI Factories and quantum machines, first come, first served (its stand is E405). If you're going, turn up early; if not, find the AI Factory nearest you and ask what a startup application involves. (European Commission)

Next edition soon,

Çelik