“Keep an eye on this” sounds like a small request. In an organisation it can mean reading confidential data, choosing what matters and interrupting someone at the right moment. An always-on agent inherits that ambiguity unless its role is made explicit.
What OpenAI launched on September 29
OpenAI announced Dots, an always-on agent powered by GPT‑6 Astra with its own cloud computer and connected tools. The initial rollout covers Pro and Business Premium in eligible markets, with an administrator-enabled beta for Enterprise, including Edu and Healthcare. OpenAI describes permissions, action reviews and visibility into the agent’s computer. There is a particularly important distinction in the announcement: proactive background research uses read-only tools and cannot send messages, edit content or control the computer. Separately authorised tasks can involve actions. OpenAI also acknowledges that agents can make mistakes. Read the Dots announcement.
Write the role before the prompt
My interpretation is that the hardest part of adopting this kind of assistant is organisational. Who owns the task? What counts as a useful finding? When should it stop?
A hypothetical reporting assistant could have a precise remit: read approved weekly sales extracts, flag changes above an agreed threshold and prepare a draft explanation. It should not change a metric definition, publish the report or message an external recipient without the appropriate approval.
That remit should be short enough to review. Specify allowed sources, the cadence, the evidence required and the destination of draft work. Name the person who resolves exceptions. “Use your judgement” is not a sufficient escalation policy when the sources disagree.
Separate observation from authority
A useful workflow has distinct stages: observe, prepare, review and publish. Moving between them should require an explicit permission boundary, especially when the task touches other people or a production system.
For an analytical draft, retain the input snapshot, query or extraction method, reporting period and glossary version. These details make an answer reproducible. Without them, a reviewer can only judge the writing.
Suppose sales are down while orders are up. The draft should show the numbers and identify what remains unknown. It should not invent a pricing explanation because that is a familiar business narrative. Evidence of a change is different from evidence of its cause.
A reviewer also needs an easy way to return the task with a specific correction. Feedback such as “wrong” is less useful than “exclude cancelled orders and use the booking date.” Save the corrected definition where the workflow can consistently use it.
Decide what deserves an interruption
Always-on work can become a new source of noise. Set a threshold for notifications and distinguish a genuine exception from an unchanged condition. Repeated alerts teach people to ignore the system.
For the reporting example, an alert could require both a material change and sufficient data completeness. If an upstream feed is delayed, the appropriate output is a data-quality warning rather than a confident trend report.
Test the workflow on quiet weeks as well as unusual ones. A system that produces a dramatic narrative every time is not exercising useful judgement.
The surrounding infrastructure matters too. NVIDIA’s new agent-safety platform addresses the question of boundaries outside the model. Delegation becomes credible when the task is defined, the evidence is visible and authority is limited to what was actually granted.
Analysis of the September 29, 2026 announcement. The reporting assistant is an illustrative workflow, not a description of a deployed client system.