Abu Dhabi Put an AI Agent in Its Top Decision Room on 17 September — the 4 Interview Questions I Regret Not Asking Agent Developers Before This Week

On Thursday 17 September the Abu Dhabi Media Office published a headline that most people in Dubai’s tech scene scrolled past: “Khaled bin Mohamed bin Zayed chairs first Abu Dhabi Executive Council meeting to use an agentic AI platform”. The National ran it the same evening, and Gulf News followed on the 18th with the decisions taken in the room. I read it as a web developer who has spent most of 2026 building agent workflows for Dubai clients, and it stopped me for a different reason than the headline. Buried in the third paragraph was a sentence that reads like an acceptance test: outputs and decisions “remain traceable and linked back to their original source papers”. That sentence is a hiring specification. This is what it says, why it matters to a private employer in Dubai, and the four interview questions I wish I had been asking agent developers before this week.

What Actually Happened on 17 September

Sheikh Khaled bin Mohamed bin Zayed Al Nahyan, Crown Prince of Abu Dhabi and Chairman of the Executive Council, chaired a regular Council session. What was new was the software in the room. The platform, named the AI Committee System, was used across the whole cycle of the meeting rather than as a demo at the end of it. Before the session, it analysed the proposals on the agenda and prepared executive summaries. During the session, it provided real-time insights and questions, fact-checked information as it was discussed, raised observations and proposed analysis-based recommendations. After the session, it produced the analytical reports and the minutes, with the recommendations and required actions summarised for follow-up.

The decisions taken in that room were not small. The Council approved the second phase of the Abu Dhabi Liveability Strategy, after a first phase that completed more than 60 projects worth AED 12 billion inside an overall programme of AED 42 billion and lifted the District Completion Index from 67% to 84% in two years. It approved the expansion of Al Ghad Nurseries, which today enrol 3,253 children across 10 sites, 80% of them children of working mothers, from 3,000 seats to 12,000 by 2029, creating more than 1,000 jobs for Emirati women. And it approved a reproductive-health programme with the Department of Health that targets an 80% IVF success rate by 2030, up from the 59% that Arabian Business reported as the current baseline. An agent sat in on all of that, and the Council said so in public.

It is also not an isolated event. Earlier in September the federal Cabinet adopted an AI advisor system to support its own decision-making, and on the same 17 September a government delegation organised by the UAE Government’s AI Office and the Dubai Centre for Artificial Intelligence was in the United States studying agentic applications for government operations. The UAE has already said it wants agentic AI running a large share of government work within two years. What changed this week is that we saw the acceptance criteria applied live, at the top of the emirate, on decisions measured in billions of dirhams.

Why a Council Meeting in Abu Dhabi Changes a Job Description in Dubai

The private employer’s instinct is to file this under government news. That is a mistake, for a simple reason: the UAE government is the largest technology buyer in the country, and its acceptance criteria travel. When the Executive Council says in public that an agent may summarise, question, fact-check and recommend only if every output is traceable to a source paper, that becomes the bar for every vendor that wants to sell an agentic product to a ministry, an authority, a free-zone regulator or a government-owned bank. Those vendors are the systems integrators, the fintechs, the consultancies and the product companies that share your talent pool in Dubai. Within a quarter, the job descriptions they publish will carry the same words.

There is a second reason. Most agent demos I have reviewed for Dubai clients this year would fail the Council’s test on the first line. They answer fluently; they do not cite. They recommend; they cannot show which paragraph of which paper the recommendation came from. They run end to end; nobody can tell you the failure rate on a held-out set, because there is no held-out set. That gap between a convincing chat window and an auditable system is exactly the gap between a junior prompt engineer and the person you actually need. The Abu Dhabi announcement makes that gap visible to buyers who could not see it before.

The 4 Interview Questions I Regret Not Asking Before This Week

I have interviewed agent developers for Dubai employers throughout 2026, and until this week my technical screen was built around frameworks and tool use. After reading the Council’s description of what the system had to do, I rewrote it. These are the four questions that are now in it, in the order I ask them.

1. “Show me an agent you shipped where every claim links to a source. Open it.”

Not “have you built RAG”. Everyone says yes to that. I want to see the citation rendered in the interface, click it, and land on the page of the document it came from. Then I ask what happens when the retriever returns nothing relevant. A strong candidate has a designed answer for that case: the agent says it cannot support the claim, or it asks for the missing document. A weak candidate has never seen the case, because their demo corpus was three PDFs and every question had an answer. The Council’s system had to work on a full agenda of live proposals; that is the scale to ask about.

2. “Show me the evaluation harness you wrote, and tell me the failure rate.”

The Council’s platform fact-checks in real time. A fact-checker with an unknown error rate is a liability, not a feature. I ask for the evaluation set the candidate built, how many items are in it, who labelled them, and what the number was the last time they ran it. If they cannot give me a number, they have never operated an agent in production; they have demonstrated one. If they give me a number that is suspiciously good, I ask what the hardest 10% of the set looks like. The best answer I have heard this year was “83% on grounded answers, 41% when the source is a scanned table, and we told the client both numbers”.

3. “Where in your system does the human decide, and how did you make that unavoidable?”

Read the announcement again: the platform prepares, questions, fact-checks and recommends. It does not decide. The Council decides. That boundary is a product decision, and it has to be built into the software rather than promised in the slide deck. I ask the candidate to draw the flow and point to the step where a person must act for anything to happen, then I ask how they stopped a busy user from clicking through it. The strong answers involve a confidence threshold below which the agent will not even propose, a diff view that shows what changed since the person last looked, and an audit log that records who approved what. The weak answer is “there is an approve button”.

4. “Run it on this Arabic PDF, on infrastructure that stays in the UAE.”

Council papers are bilingual, and a large share of the source documents any UAE government or regulated-sector agent will touch are Arabic, scanned, or both. Data residency is not negotiable for most of those buyers. So the last question is practical: I hand the candidate a real Arabic document and ask how their pipeline would ground an answer in it, which OCR and embedding choices they would make, and where the inference would run. A candidate who has only ever deployed on a US-region endpoint with an English-only retriever will tell you so in the first thirty seconds, which is exactly what you need to know. The seven-step guide we wrote after the government’s 50% agentic pledge covers the residency side in more depth.

3 Things to Change in Your Next Agent Developer Requisition

You do not need to rewrite your hiring process because of one meeting in Abu Dhabi. You need to change three lines in the next job description, and the way you run the technical round.

• Replace “experience with LLM frameworks” with “has shipped a source-grounded agent and can show it”. Frameworks change every quarter; the discipline of grounding does not. Ask for the link in the application form and disqualify on its absence.
• Add a line on evaluation, with a number. “You will own an evaluation set of at least 200 labelled cases and report its pass rate before every release.” Candidates who have done this will recognise it; candidates who have not will self-select out, which saves everyone a week.
• State the residency and language constraints in the first paragraph. If your customers are UAE government entities, banks or free-zone authorities, say that inference runs in-country and that the corpus is bilingual. It is the cheapest filter you have, and it moves the Arabic-capable candidates, who are rarer than you think, to the top of the pile.

For the technical round itself, drop the take-home that asks for a chatbot. Give the candidate a folder of six real documents, two of them in Arabic and one a scanned table, and ninety minutes with your senior engineer on a call. Ask for grounded answers to five questions and a one-paragraph note on what they would evaluate before shipping. You will learn more in that session than in any portfolio review. If you are building a product for the public sector specifically, our note on building government services applications in the UAE lists the procurement constraints that shape the stack.

If You Also Hire in Singapore

The same week Abu Dhabi put an agent in its Council chamber, Singapore’s market was absorbing three frontier AI employers landing inside 72 hours, and the competition for exactly this profile intensified there too. Our Singapore team’s analysis of why OpenAI, Cognition and Anthropic chose Singapore as their regional hub and their seven-step playbook for competing with Google Cloud for engineers apply almost line for line to a Dubai employer competing with the government’s own vendors. If you run teams in both cities, the four interview questions above are the same in both; only the language of the test documents changes.

FAQ — The Abu Dhabi Executive Council’s Agentic AI Meeting and Dubai Hiring