OpenAI announced dots on September 29, 2026: agents that stay on, keep their own cloud computer, and connect out to the apps a business already runs on. The announcement says they reach "over 4,000 apps" through plugins and that you can message them in Slack and Teams.
That number is the interesting part, and also the risk. An agent with thousands of reachable apps and a standing mandate to act is a different thing from a chatbot you ask for a draft. The first week of configuration is what determines which of the two you get.
The announcement describes dots as always-on agents with their own cloud computer, working continuously through connected apps. Availability is Pro and Business Premium users in eligible markets, with an Enterprise beta that a workspace admin has to enable.
The controls it documents are the part to read twice:
The announcement also says the honest thing: dots can still make mistakes, so review consequential work. That sentence is the whole reason the setup below matters.
The common mistake is connecting the tools first and hoping a useful pattern emerges. Work the other way.
Pick one recurring task that has a clear start, a clear end, and a result you can check. The announcement's own examples are instructive because each one has a finish line: a bug appears in Slack and gets investigated, a design arrives and becomes a working app, a planning cycle stays in sync. Procurement, invoice processing, email marketing, customer support and contracting are named as further areas.
Then ask one question about that task: if it went wrong quietly, how would you find out? If there is no answer, that task is not the place to start, because you have no way to catch the failure.
Built-in rules handle the obvious cases. Custom Rules exist for the ones your business cares about, and the useful test is reversibility rather than importance. Anything hard to reverse goes behind approval, even if it looks minor: sending a message to a customer, changing a record that other systems read, paying or refunding, deleting, publishing.
OpenAI's guidance for agents that operate software puts the same boundary in operational terms, keeping a person in the loop for purchases, authenticated flows and destructive actions, and warning that content the agent reads cannot grant it permission. A dot reading a supplier email is not authorisation to pay the invoice.
The comparison worth keeping in mind is that this is now a standard shape rather than an OpenAI quirk. Google's agent announcement of October 8 bundles identity, permissions, sandboxing and spending limits, and Microsoft describes Autopilot as persistent and proactive, with "permissions, audit and governance behind it", while making clear you set the objective and the boundaries. Every vendor hands you the controls. None of them writes your rule.
That rule layer is the same one described in Google ships agent controls, and it is why bounded permissions and human approval is a design decision rather than a setting.
A useful constraint for the first month: one dot, one job, one channel. Not because the product cannot do more, but because a single dot with a single measurable job is the only configuration where you can tell a good result from a plausible one.
Refuse the temptation to let the dot own a process end to end in the first week. Give it the part of the process you can check, keep the rest manual, and move the boundary outward as you build evidence. That is the difference between adopting an agent and discovering one running your business.
If the work involves money, records or customers, the review points follow the pattern in read only versus read write access and the guardrail argument in why your AI agents need a human guardrail. If you are still working out whether this is the right role to delegate at all, AI agents for small business is the earlier question.
Four signals say the setup is working, and they are all about behaviour rather than output.
An always-on agent earns its place where work arrives unpredictably and needs a response. If the task runs on a schedule with the same steps each time, a plain automation is cheaper and easier to reason about. If the task needs judgement on every single instance and no step is repeatable, the result will be a queue of approval prompts and a person doing the work anyway, which is the pattern described in 66% use AI, 9% embed it. Persistent agents fit the middle: recurring work with variable inputs and checkable outcomes.
No. The announcement puts the rollout on Pro and Business Premium in eligible markets, with Enterprise, Edu and Healthcare able to use a beta once a workspace admin enables it.
Only where the first job needs them. Each connection widens what the agent can reach, and the permission you set is the only thing standing between a mistake and a customer.
They cover the obvious cases. Custom Rules are where your business's own line between routine and consequential gets written, and that line is specific to you.
That is what the review step is for. The announcement says to review consequential work, which only means something if a named person is doing that review and can stop the work while they check.
Maxpertise builds and runs AI teams that take recurring work off your team, in a managed lane where we host and run the team, or an embedded lane where AI engineers work inside your existing team. Both lanes keep the two human gates: a person approves the plan before the work starts, and a person reviews the result before it ships.
Persistent agents make that structure more valuable, not less. The setup decisions above, the first job, the app scope, the rules and the review owner, are the same decisions we start from in an AI back office, and they belong in a written spec you can read rather than in a configuration screen only one person understands. If you want the method, start at how it works; the engagement options are on pricing, and contact with the one task you want to start with is enough to scope it.
[1] OpenAI, Introducing dots, September 29, 2026. Always-on agents, plugin reach across 4,000 apps, Slack and Teams messaging, app controls, built-in rules and Custom Rules, availability by plan, Agent 365 integration, and the instruction to review consequential work. Read through the research agent: openai.com returns HTTP 403 to direct fetches from this machine, including its homepage.
[2] Microsoft, Introducing the new Copilot with Home, Code and Autopilot, September 25, 2026. Autopilot as a persistent and proactive agent, the objective and boundary framing, permissions, audit and governance, and the plugin registry with central approval.
[3] Google Cloud, Welcome to Gemini at Work 2026: Introducing the Gemini agent, October 8, 2026.
[4] OpenAI, Computer use integration: handle user confirmation and consent, accessed October 10, 2026.
Maxpertise is an AI-native engineering company. We embed native AI engineers inside your team, live in about 10 days. Please enable JavaScript to view the site, or email contact@maxpertise.net.