This fortnight the rulebook for AI agents started getting written. The UAE published a design guide for AI-run public services, and one line in it is worth pinning to the wall: human leads, AI enables. In open source, the two big standards for getting agents to talk to each other moved under the same neutral roof, so it's getting easier to swap tools instead of being stuck with one. In Riyadh, Saudi Arabia signed a deal worth hundreds of millions of euros to run frontier AI on its own infrastructure, in Arabic. So the plumbing and the rules are coming along nicely. The part nobody has written down is the people. In a new survey, 43% of workers said their AI skills are already behind what the job needs, and more than a third are learning in their own time to keep up. Inside: the news, how to run an AI rollout that works in Arabic, how to buy AI you can still walk away from, and our new playbook on deciding what your agents may decide.
The two protocols that let AI agents talk to each other now sit under one neutral roof. That's the difference between tools you can swap and tools you're stuck with.
250+ — members of the Agentic AI Foundation, up from fewer than 40 at its launch in December 2025 (Axios, 17 August 2026)
If you run more than one AI agent, you already have a plumbing problem. An agent built on one vendor's platform doesn't naturally talk to an agent built on another's. Until now, you paid someone to hand-build a connector for every pair. Multiply that across a few dozen workflows and the integration bill quietly becomes the biggest line in the project.
That's what these two protocols are for. MCP handles how an AI application reaches your tools and data. A2A handles how one agent talks to another. Putting both under the same neutral foundation, with the major cloud and model providers all in the room, makes it much more likely they stay compatible rather than drifting into rival camps.
The practical payoff is bargaining room. If your agents speak common protocols, you can move a workload to a cheaper or faster provider next year without rebuilding it. Ask your vendors which protocols they support, and get the answer in writing before you sign.
Why it matters: Most large employers will end up running agents from several vendors. Common standards mean you stop paying to connect every pair by hand — and you keep the option to move.
Source: Axios
Four in ten workers say their AI skills are behind what the job needs. More than a third are training in their own time to catch up.
43% — of workers say their AI skills are behind what they need to stay competitive (Resume Now, The AI Skills Race Report, survey of 1,000+ employed workers, 25 August 2026)
This is not a training-budget story. It's a retention story. When people believe AI skill quietly decides who gets promoted, and their employer hasn't said a word about how it will help them, they do one of two things. They learn in secret, badly and slowly. Or they start looking at employers who have said something.
The hiring market backs up the anxiety. ManpowerGroup surveyed more than 39,000 employers across 41 countries, including the UAE, and found that for the first time AI skills are the hardest of all skills to hire — AI model and application development at 20% and AI literacy at 19%, both ahead of engineering. Seventy-two percent of employers report trouble filling roles, and the most common response is upskilling their own people.
The fix is cheaper than the anxiety suggests. Name the two or three AI skills that actually matter in each team. Fund the hours during work time, not evenings. Say out loud that nobody is expected to figure this out alone. That last sentence does more for retention than another tool licence.
Why it matters: If you haven't told your team how the company will help them get good at AI, they've assumed the answer is "it won't" — and some of them are acting on it.
Source: Resume Now / CPA Practice Advisor
HUMAIN and Mistral signed a deal worth hundreds of millions of euros — to run advanced AI on Saudi infrastructure, and to make it work properly in Arabic.
Hundreds of millions — of euros — the size of the HUMAIN–Mistral collaboration, per the joint announcement (HUMAIN and Mistral joint announcement, Riyadh, 24 August 2026)
HUMAIN, the PIF-owned AI company, and France's Mistral announced a collaboration covering infrastructure, model development and deployment across Saudi Arabia and the wider region. They'll start with cybersecurity and voice, and build frontier models that perform strongly in Arabic. Mistral will look at running on HUMAIN's data centres. The two plan to sell into regulated industries together — financial services, telecoms, manufacturing and the public sector.
Why it matters: The announcement spells out what sovereign AI means: your data stays inside boundaries you set, you can own and adapt the model weights, and training and inference run in a jurisdiction you choose. In the Gulf that is turning from a preference into a procurement requirement.
What to do: Add three lines to your next AI vendor questionnaire. Where does our data physically sit? If we leave, do we keep the model weights we paid to tune? And does it work in Arabic to the same standard as English?
Source: HUMAIN / Mistral
There's now a published rulebook for AI-run government services. The best line in it: human leads, AI enables.
50% — of UAE government operations, services and tasks to run on agentic AI within two years — now with a published design rulebook for how (Global Government Forum, reporting the National Committee for the Agentic AI Project and WAM, 26 August 2026)
We covered the UAE's target last edition — half of federal government operations running on AI agents within two years. This is the how. The National Committee for the Agentic AI Project released a unified design guide at a workshop with more than 100 federal officials. Its principles: never ask someone for information they've already given you, keep them informed at every step, protect privacy and security, unify payments, be clear about the final outcome, and always let them object or correct. It plans deliberately for the moments a human has to step in.
Why it matters: This is decision rights, written down — what the agent handles alone, what it escalates, and what a person can always appeal. Most private-sector agent projects have none of this on paper, which is exactly why they stall somewhere between pilot and production.
What to do: Print the six principles. Put your own service names beside them. Whichever ones your current agent pilot would fail is your next fortnight's work.
Source: Global Government Forum
Look at almost any AI rollout in the UAE or Saudi and you'll find the same shape. The policy is in English, the prompts are in English, the outputs are in English, and Arabic gets bolted on at the end by whoever is free. That last step is where the mistakes happen and where the deadline slips.
If a process only works in English, it only half works.
AI tools are improving faster than any procurement cycle can keep up with. The tool that's best today probably won't be in eighteen months. So the question to ask in the room isn't only "is this the best one" — it's "what does it cost us to change our mind".
Flexibility is much cheaper to buy at the start than to retrofit later.
Playbook. Six steps to write down what your AI agents may decide on their own, what has to come back to a person, and who answers for it. Built for C-suite and directors in the UAE and Saudi Arabia — the output is one page your team can act on this week.
Hyrra.ai. Hyrra.ai puts AI agents to work on the CV pile — it does the first-pass screening and cuts the time it takes to build a shortlist, with a recruiter in control of every shortlist and every rejection. That control is the point: the agent handles the volume, a person still owns the decision.