Verdict: partly held. The emphasis on access boundaries and review is reinforced, but the earlier note understated how much interruption and auditability must extend beyond the agent’s own controls.
Original analysis: Dots makes permission design part of the agent, not an afterthought, published 2026-09-30 (6 weeks ago).
What we said
New Runtime’s earlier coverage argued that dots makes permissions part of delegation: users choose app access, review work, and can redirect or pause the agent. It said builders should treat permission UX and auditability as core workflow features, test whether people understand an agent’s authority, and avoid planning around unverified connector claims. The analysis focused on the practical safety of real delegated work, rather than demo performance or connector count.
What happened since
- New Runtime’s October 5 coverage returned to the same design question, arguing that access boundaries matter more than the number of integrations. It described app selection, Custom Rules, review, redirection, and pausing as parts of delegation, reinforcing the earlier emphasis rather than shifting toward connector counts. That is a useful sign of continuity in the newsroom’s analysis, not independent proof that the controls work well in practice [1]
- A study of software practitioners found that permission decisions partly depend on what people can observe, while behind-the-scenes activity creates uncertainty. Its recommendations include making consequential actions easier to review, separating what an agent is allowed to do from what a user intended, explaining reversibility, and not treating repeated approvals as lasting consent. This supports the earlier call for understandable previews, wh [2]
- OpenAI’s workspace guidance documents app-level access and approval controls, distinct controls for browser, network, and computer access, and administrator pause/resume and audit events. But it also says pausing a dot does not undo completed actions and may not stop separately delegated or scheduled tasks; administrators may not be able to verify that all work has stopped. The practical boundary of a “pause” is therefore broader than the agent t [3]
- In an October 13 update, OpenAI said approval buttons should no longer be clipped or hidden, safety banners were more stable, and failed activity views would retry loading. These concrete changes make approvals more visible and review more dependable, directly addressing small but consequential interaction problems. They do not establish that access is sufficiently narrow or that audit coverage reaches every related task. [4]
What we missed
The original analysis treated pausing, review, and auditability mainly as features of the agent interaction. It did not adequately distinguish stopping the agent from stopping other delegated or scheduled work, nor stopping future activity from reversing actions already completed. The workspace guidance also leaves a practical verification problem: an administrator may not be able to confirm that everything has stopped. That is material, because a control that looks decisive in the interface may not define the boundary of the full workflow.
Forecast check
- We expect OpenAI’s public dots materials to clarify supported integrations or availability before the broad connector claims become dependable planning assumptions. (65%, due 2026-12-30): unclear. Workspace guidance documents access and approval controls, but the findings do not establish a verified supported-integration list or resolve availability. The forecast is not due until December 30, 2026.
- We expect the next wave of discussion about dots to focus on approval friction and user control, not only task demonstrations. (60%, due 2027-03-30): unclear. The practitioner study and product update both raise review, approval visibility, and user control, an early sign consistent with the forecast. The evidence does not establish that the next broad wave of dots discussion will focus on friction and control; the forecast is due March 30, 2027.
- We expect persistent-agent evaluations to include permission scope and interruptibility alongside task completion. (55%, due 2027-09-30): unclear. The study addresses permission decisions and recommends attention to reversibility, but it is not evidence that persistent-agent evaluations now include permission scope and interruptibility alongside task completion. The forecast is due September 30, 2027.
What to do now
Builders should continue to map each delegated workflow into separate read, write, approval, and stop boundaries, but make the map cover the entire chain of work—not just the visible agent session. Identify whether browser, network, computer, scheduled, or separately delegated activity is involved; decide how a person can verify that each part has stopped; and make clear which completed actions cannot be undone. Show consequential actions in a reviewable form, distinguish the agent’s permitted authority from the user’s intent, and avoid treating repeated approvals as durable consent. Buyers should test these boundaries in realistic workflows, including a pause during active work, rather than accepting the presence of a pause button or an audit log as proof of control.
Next check
Test whether an administrator can reliably verify that all agent-related and separately delegated or scheduled work has stopped after a pause, and identify what completed actions remain irreversible.