On 31 August 2026, Salesforce published its Agentic Enterprise Index. The finding that travelled fastest was the count: companies now operate an average of 13 active AI agents, against roughly 5 a year earlier. Nearly a tripling in twelve months.
That is the number everyone quoted. It is not the number that matters.
Buried in the same report is a second figure: the time required to build an agent has fallen by 53%. That is the finding with consequences, and it explains the first one. Agents did not triple because companies became three times more ambitious. They tripled because building one stopped being a project and became a task.
We run technical vetting for UAE employers. Over the last quarter we conducted 31 engineering interviews in Dubai for roles touching AI agents in production. Here is what this shift has already done to what we screen for — and what most job descriptions in this market are still getting wrong.
Why cheap agents create an expensive problem
When something becomes ten times cheaper to produce, you do not get the same amount of it for less money. You get far more of it, and the cost moves somewhere else entirely.
An organisation running three agents can hold the whole picture in one person’s head. Someone knows what each agent touches, what permissions it holds, and roughly what it costs. At thirteen, nobody knows. The agents were built by different teams, at different times, against different data, with permissions granted for reasons no longer documented.
The engineering problem stops being can we build this and becomes can we operate a fleet of these without one of them quietly doing something expensive. Those are not the same skill, and they are frequently not the same person.
The reflex that changes everything
Stop asking candidates what they have built with agents. Ask what has gone wrong with one, and how they found out. The first question rewards anyone who has followed a tutorial. The second can only be answered by someone who has operated an agent that had access to something real. In our 31 Dubai interviews, this single substitution changed the ranking of the shortlist entirely.
Expert view 1: the market is still hiring for the phase that just ended
Read job specifications currently open in Dubai and Abu Dhabi for agent-related roles. The overwhelming majority describe construction: build agents, integrate tools, design prompts, connect data sources.
That is hiring for 2025. If build time has fallen by half, construction is precisely the part of the work that is deflating in value. The scarce capability is everything that happens after an agent exists and before it causes an incident.
This mismatch is very visible in interviews. Candidates arrive well prepared to discuss architecture and orchestration frameworks, and are frequently unable to answer basic operational questions about systems they have themselves deployed. That is not a criticism of the candidates — it is a criticism of what the market has been asking them to optimise for.
The 4 skills that became disqualifying if absent
These are not aspirational. Each one caused at least three rejections in our last 31 interviews.
1. Permission scoping
Can the candidate explain how they decide what an agent is allowed to touch, and how that decision is reviewed later? The weak answer describes giving the agent the access it needs to work. The strong answer starts from the smallest possible scope and describes an approval trail. With thirteen agents, over-permissioning is not a theoretical risk — it is the default state unless someone actively fights it.
2. Agent-level observability
This is where most candidates fail. The question we use: an outcome three days ago was wrong, and four agents touched that record — how do you find which one did it? Nineteen of thirty-one could not attempt an answer. Standard application logging does not solve this, because every agent action looks like a successful API call.
3. Conflict handling
What happens when two agents act on the same record within the same second? Anyone who has run more than a handful in production has met this and will describe locking, sequencing, or an explicit ownership model. Anyone who has not will say it has never come up — which is the honest answer, and also the disqualifying one for a fleet role.
4. Cost attribution
Can they tell you what a single agent costs to run per month, and per business transaction? At five agents nobody asks. At thirteen, finance asks, and the engineer who cannot answer becomes a problem for their manager. This is now the fastest way to differentiate two otherwise similar candidates.
Hiring for agent operations in the UAE?
We screen against these four skills specifically, with written scenarios rather than framework trivia, and pass you candidates who have operated agent fleets rather than demonstrated them. You get the interview transcripts, not just a CV.
Let’s talk — brief us on the roleExpert view 2: the governance threshold is around 3 agents, not 13
The average of thirteen makes this look like a large-enterprise concern. In the UAE market, where a great deal of the interesting work happens in companies of thirty to two hundred people, that reading would be a mistake.
From what we observe, the point at which things stop being manageable informally is roughly three agents. That is where no single person can still recite what each one touches. Most Dubai companies that deploy their first agent cross that line within a quarter, because the second and third are now cheap — which is exactly what the 53% figure predicts.
So the practical guidance for a smaller UAE employer is not wait until you are big enough to worry. It is: whoever builds your first agent should be someone who can also answer the four questions above, because they will be building your fourth one before the year is out.
Expert view 3: this widens the gap between the UAE and offshore delivery
There is a structural consequence for how UAE companies source engineering work. Building agents is highly outsourceable: the specification is clear, the deliverable is testable, and the work is bounded. Operating a fleet of agents is not, because it requires standing access to production systems, ongoing context about the business, and the authority to intervene.
We expect this to push the market toward a split that is already visible in our engagements: construction distributed to wherever it is most efficient, operations held close. For UAE employers this argues for hiring or retaining a small number of people with genuine fleet-operations capability, rather than a large team of agent builders.
If you are comparing how this is playing out elsewhere in the region, HireDeveloper.sg tracks the same shift in Singapore, where agentic adoption started earlier and the operations-versus-construction split is further advanced. JapanDev covers the Tokyo market, where the governance question has been framed around data provenance rather than permissions — a useful contrast when you are writing your own controls.
The interesting number in that report was never thirteen. It was fifty-three per cent. Cheap creation is what turns a manageable set of systems into a fleet nobody has mapped. — William, HireDeveloper.ae
What to change in your next job specification
Three edits, all of which take ten minutes:
- Replace « experience building AI agents » with « experience operating AI agents in production ». Different candidates will apply. That is the point.
- Add one scenario question to the screening stage, in writing, before any call. Ours is the three-days-later attribution question. It removes roughly half the pipeline before you spend interview time on it.
- Ask for a cost figure. Any candidate who has run an agent against a real budget will have one. It is a fifteen-second filter with an unusually high signal.
For the build-side decisions that sit upstream of all this, our guides on building an AI agent platform in the UAE and building a CRM system in the UAE set out the scoping work that makes these roles definable in the first place.
Start with one question on your next call
Ask your next candidate how they would identify which of four agents caused a wrong outcome three days ago. Nineteen of our last thirty-one could not attempt it. It takes one minute and reorders your shortlist.
Let’s talk — see vetted candidatesFAQ
What did the Salesforce Agentic Enterprise Index actually report?
Published on 31 August 2026, it found enterprises now operate an average of 13 active AI agents against roughly 5 a year earlier, and that build time per agent fell 53%. The count is the headline; the build-time collapse is the operationally significant finding, because it moves the constraint from building to governing.
Does this mean UAE employers should hire fewer engineers?
No — the data points the other way. Faster creation increases the number of production systems requiring monitoring, permission management and failure handling. In our 31 Dubai interviews, the roles under most pressure were not agent builders but the people keeping fleets running safely. Demand shifted rather than shrank.
How do you test for multi-agent experience in an interview?
Ask what happens when two agents act on the same record in the same second, and how they would identify which agent caused a bad outcome three days later. Neither is answerable from tutorials, and both are trivial for anyone who has operated agents in production. Nineteen of our thirty-one candidates could not attempt the second.
Is this relevant to smaller UAE companies?
It becomes relevant earlier than most founders expect. The governance problem starts at around three agents — the point where nobody can hold the full permission map in their head — not at thirteen. Smaller Dubai companies typically cross that line within a quarter of their first deployment.