Most of the DevDay coverage on 29 September led with the same framing: OpenAI now has its own answer to the persistent-assistant products its rivals shipped earlier in the year. That is true and it is not very interesting. What held my attention was a single sentence about rollout mechanics — Enterprise, Edu and Healthcare workspaces receive dots once an administrator enables them. One toggle, owned by whoever administers your workspace, that grants a continuously running process its own browser and its own credentials. By the end of that week I had rewritten four Dubai job specifications we were about to publish.
What OpenAI Actually Shipped on 29 September
The keynote ran on 29 September 2026 and carried more than twenty announcements. Four of them matter if you are responsible for an engineering team.
Dots. Always-on agents that live inside ChatGPT. Each runs on GPT-6 Astra, and critically each gets its own cloud computer and browser. They work toward a goal continuously rather than only during a conversation, and they reach more than 4,000 applications through plugins. Availability started with Pro and Business Premium, with Enterprise, Edu and Healthcare workspaces enabled by an administrator.
GPT-6.1 Sol. An upgrade pitched explicitly at agentic coding, computer use and professional work, which OpenAI says approaches Astra-level intelligence at roughly twenty percent of Astra token pricing.
Ultrafast. A premium speed tier for GPT-6 Astra on Pro 500 and Enterprise plans, priced at about six times standard API pricing, claiming up to eight times faster responses in Codex and up to six times faster in the API.
The model that did not ship. OpenAI said it would not release GPT-6.1 Astra, its newest frontier model, citing security concerns. A vendor holding back its best model is a data point about the maturity of this category, and it belongs in your risk assessment rather than in your disappointment.
Supporting launches filled in the surface area: Codex now runs in the cloud with voice command-line support, ChatGPT Space provides a shared team workspace, Pages is a document editor built for human and agent collaboration, and Sign In with ChatGPT arrived across sixteen partner tools. Details are in the DevDay announcement thread on the OpenAI developer forum and in BGR’s rundown of every announcement.
Our expert take #1
The announcement that changes hiring is not the model. It is its own cloud computer and browser. The moment an agent holds a browser session, it inherits every credential that session can reach, and it does so without a human watching the screen. That is not a prompt problem and no amount of prompt review will contain it. It is the same class of problem as a long-lived service account, and the people who already know how to contain those are sitting in your platform and infrastructure teams, not in your AI channel.
The Admin Toggle Is the Whole Story
Read the rollout mechanics again: Enterprise, Edu and Healthcare workspaces get dots once an administrator switches them on.
In a Dubai company of two hundred people, who is that administrator? In my experience it is an IT manager or an operations lead. It is almost never the person who would be asked to review a design document for a long-running autonomous process with network egress. So a decision with the risk profile of deploying a new production service gets made through an interface that looks like a settings page.
I do not think that is OpenAI being careless. Every vendor ships enterprise features behind an admin switch. The point is narrower and it is about your side of the line: the control that gates this capability sits outside your engineering review process by default, and nobody will move it there unless somebody owns the job.
That ownership gap is what I went looking for in our open specifications, and it is what I did not find in any of them.
Why Cheaper Tokens Will Raise Your Bill
GPT-6.1 Sol at roughly a fifth of Astra pricing reads like a budget win. It is not, and the reason is behavioural rather than technical.
Cost was the main thing keeping agent usage narrow at most companies I work with. Teams ran agents on the three workflows that clearly justified the spend, and the rest stayed on the backlog. Remove four fifths of the unit cost and those backlog items become defensible overnight. Volume does not rise by twenty percent to match the saving. It rises by whatever the organisation can think of, which is a much larger number.
Then add the other half of the announcement. Ultrafast costs about six times standard API pricing. The moment one team discovers that a latency-sensitive workflow feels better on Ultrafast, your blended rate stops resembling the headline figure in anybody’s business case.
The teams that come through this without an unpleasant quarter are the ones that treat spend per agent run as a hard engineering limit, enforced in code, before they widen adoption. Not a dashboard. A ceiling that stops the run.
Our expert take #2
Ask any candidate who claims agent experience one question: what stops a runaway agent run in the system you built, and who gets paged when it triggers? Someone who has genuinely operated agents answers in under a minute with a specific mechanism — a token ceiling, a step limit, a wall-clock timeout, a circuit breaker on a tool call. Someone who has only built demos talks about prompt design and evaluation suites. Both answers sound informed. Only one of them has been on call.
Writing an agent-facing role this quarter?
We screen UAE engineers on containment design, credential scoping and budget enforcement before they reach your calendar — so the first call is not spent finding out they have only ever run agents in a notebook.
Discutons-en — talk to our Dubai teamThe Four Dubai Specifications I Rewrote
These were live specifications for Dubai clients that week. Here is what changed in each, concretely.
1. From “AI Engineer” to “Agent Runtime Engineer”
The old version asked for LLM experience, RAG pipelines and a vector database. In the UAE market that description matches several hundred people and therefore sorts nobody. The rewrite asks for experience running processes that continue without a human watching, and names the artefacts: step limits, wall-clock timeouts, idempotent tool calls, replayable logs. The screening question is now: describe a long-running job you owned that failed badly, how you found out, and what you added so it could not happen the same way twice. Note that the question never mentions AI. That is deliberate, because the skill predates it. Our breakdown of agent operations hiring after the Mac fleet computer-use launch covers how this profile first appeared.
2. From “Security Engineer” to “Agent Identity and Credential Engineer”
An agent with its own browser is an identity problem wearing a product name. The rewritten specification screens for short-lived credential design, scoped service identity and egress control — and specifically for the judgement to say no to a convenient but over-broad token. I added one scenario to the technical screen: a dot needs to read invoices from a shared mailbox and write summaries to a finance system. Design the credential. Tell me what it must not be able to reach. Strong candidates immediately ask who approves the scope and how it gets revoked.
3. From “Platform Engineer” to “Agent Cost and Capacity Engineer”
This one is new, and I expect to see it on a lot of org charts within two quarters. With Sol cheap and Ultrafast expensive, somebody has to own model routing by workload class, per-run spend ceilings, and the capacity model that predicts next month. That is FinOps with an inference vocabulary. The candidate pool is deeper and materially less expensive than the AI-specialist pool most companies are recruiting from by reflex.
4. From “IT Administrator” to a Spec With an Engineering Reviewer Attached
The fourth change was not a new role. It was adding a line to an existing administrator specification stating that enabling autonomous agent features in any workspace requires engineering sign-off, and naming who gives it. That sentence costs nothing, takes a minute to write, and is the single highest-leverage change I made that week. It closes the toggle gap described above before an incident makes the case for you.
The Case Against Overreacting
I want to be fair to the other reading of this, because I have heard it from people whose judgement I respect.
Dots are, for now, a consumer-shaped feature that reached Pro and Business Premium first. Enterprise adoption needs an administrator to act. Plenty of Dubai companies will not enable it this quarter, and some will never let an autonomous browser session near their systems. If that is you, none of this is urgent, and rewriting four specifications in a week would be theatre.
There is also a real argument that this category is less mature than the marketing suggests, and OpenAI supplied the evidence itself by declining to release GPT-6.1 Astra over security concerns. A vendor that holds back its best model is telling you something worth hearing.
So the honest version is conditional. If your workspace administrator could enable this tomorrow and no engineer would be consulted, you have an ownership gap and it is worth one afternoon. If your change process already catches this, you have time, and you should spend it on the cost controls rather than the security ones — because the budget surprise is the near-certain outcome and the security incident is only a possible one.
Our expert take #3
Dubai employers have a genuine structural advantage here that I do not see discussed. Teams in the UAE are typically smaller, newer and less burdened by legacy change-management than their counterparts in London or Frankfurt, which means the governance for autonomous agents can be designed once, correctly, instead of retrofitted across twenty years of accumulated process. That advantage has a short shelf life. It lasts exactly as long as it takes for the first team to enable the toggle without telling anyone.
What I Would Do This Week
- Find out who can flip the toggle. Ask your workspace administrator directly whether autonomous agent features can be enabled today and whether anyone would be consulted. The answer takes five minutes to get and tells you whether the rest of this list is urgent.
- Write the spend ceiling before the pilot, not after. A per-run token limit enforced in code. With Sol at roughly a fifth of Astra pricing, usage will expand faster than your finance team expects.
- Add one question to every technical screen for infrastructure and platform roles: what stops a runaway unattended process in a system you built. It costs four minutes and it sorts the market.
- Resist creating an AI team. Three of the four changes above belong to engineers you already employ. Giving one of them explicit ownership of agent runtime for a quarter beats recruiting a specialist you cannot yet write a specification for.
- Separate the credential question from the model question. They get conflated constantly, and only one of them can cost you a customer.
If you are weighing whether this profile belongs in-house or with a partner, our guide to hiring AI browser automation engineers in Dubai covers the screening sequence in more depth, and our build guides show where these controls sit in a real system. If the same roles are on your Singapore roadmap, the regulatory backdrop differs enough to be worth reading separately — our colleagues cover how to evaluate agent security skills when hiring in Singapore, and their note on the Agents API beta and APAC engineering hiring tracks the same shift from the other side of the region.
Hiring into a category that did not exist in June?
We source engineers across the UAE and pre-screen them on containment design, credential scoping and cost control — and we will tell you when the role you have written is really a training plan for someone already on your team.
Discutons-en — brief our Dubai teamFrequently Asked Questions
What did OpenAI actually announce at DevDay on 29 September 2026?
The headline launch was dots: always-on agents that live inside ChatGPT, run on the GPT-6 Astra model, and each receive their own cloud computer and browser so they can keep working toward a goal continuously rather than only while a human is in the conversation. Dots can reach more than 4,000 applications through plugins. They rolled out first to Pro and Business Premium subscribers, with Enterprise, Edu and Healthcare workspaces able to switch them on through an administrator. Alongside dots, OpenAI announced GPT-6.1 Sol, which it positions as stronger at agentic coding and computer use while costing roughly twenty percent of Astra token pricing, and an Ultrafast speed tier priced at about six times standard API rates. OpenAI also said it would not release GPT-6.1 Astra, its newest frontier model, citing security concerns.
Why does an always-on agent change a job specification rather than just a tool list?
Because the failure mode changes shape. A chat assistant fails in front of the person who prompted it, and that person notices immediately. An always-on agent with its own browser and credentials fails at three in the morning, several steps into a plan, against a system nobody is watching. The work that makes that safe is not prompt engineering. It is budget enforcement, credential scoping, idempotency, replayable audit logs and a kill switch that someone has actually tested. Those are platform and reliability engineering skills, so the specification should be screening for distributed systems judgement rather than for familiarity with a model vendor.
Is the cheaper GPT-6.1 Sol model good news or bad news for a Dubai engineering budget?
Both, and the order matters. A model that is close to frontier quality at roughly a fifth of the token price removes cost as the reason teams kept agent usage small. In practice that means volume rises faster than the per-token saving falls, and total spend goes up rather than down in the first quarter. The teams that stay solvent are the ones that put a hard ceiling on spend per agent run before they expand usage, which again is an engineering control rather than a procurement one. Budget the control work in the same sprint as the adoption, not after the first invoice.
Should a Dubai employer hire a dedicated agent operations engineer or train the existing team?
For most teams below roughly thirty engineers, train first and hire second. The skills that matter are already adjacent to what a competent site reliability or platform engineer does, and the fastest route is to give one existing engineer explicit ownership of agent runtime, budgets and credentials for a quarter. Hire externally when you cross the point where agents touch production data or customer-facing money movement, because at that point you want someone who has already seen an autonomous process fail badly and can design the containment before it happens rather than after.
Panos Petropoulos
Web Development Expert at HireDeveloper.ae. Reviews engineering job specifications and technical screens for UAE platform and AI teams.