The Dots launch is being read as a capability story — agents that run 24/7 on their own cloud computer, powered by GPT-6 Astra, wired into 4,000-plus apps. That’s the least interesting part. Keeping an agent running was never the bottleneck. Knowing whether it did the right thing while you weren’t watching is.

What actually ships here is an oversight layer. Dots run any account-affecting or information-sharing action through an auto-review that checks it against your instructions, Custom Rules, and a safety policy before it commits. And the feature OpenAI keeps highlighting — that you can open the Dot’s computer and watch it work mid-task instead of waiting for a finished answer — is a quiet admission: continuous agents break the request/response contract that made LLM apps auditable. When work happens in the background across many steps, your eval harness can’t score a single turn anymore. You’re evaluating a process, not a response.

That reframes the enterprise problem. In production agentic work the failure modes worth losing sleep over aren’t “the model can’t do it” — they’re silent state mutation, an action taken on stale context, a tool call that looked fine in isolation and wrong in sequence. Always-on makes every one of those harder to catch, because there’s no human in the loop at the moment of the mistake. The real engineering isn’t the agent; it’s the observability, the rollback story, and whether Custom Rules are expressive enough to encode intent an agent can’t infer.

The public reaction tracks this exactly. TechCrunch and VentureBeat covered it straight, with VentureBeat flagging that an agent acting on authenticated business software can quietly alter records; Traictory was sharper, calling the safety story a launch-film design document with no independent testing and reading the metered tiers as a pricing ladder built in public. None of the doubt is about whether Dots are capable — it’s whether the guardrails hold, which is the same question I’d ask before pointing one at anything real. The HN discussion settles in the same place.

tags: [ agentic-ai ] [ llm-ops ] [ industry ]