What your always-on agent wakes you for

OpenAI launched Dots this week: always-on AI agents equipped with their own computer, browser, apps, context, and memory. You can give a Dot access to your tools, teach it your working style, and it will continue working while you’re away — researching, monitoring, and preparing — returning to you only when it needs a decision.

The last clause is crucial. The announcement gained significant attention, amplified by a "hidden launch video" meme post, despite some initial skepticism labeling it as "a glorified Codex session." However, the interesting design question lies not in what the agent does, but in what prompts it to wake you up.

01The hour and a half

Skeptics should take note of what quietly became standard this month. A founder tasked Opus 5.5 with designing a working LEGO replica of the Ford Model T, complete with instructions and a list of pieces to buy. The model ran for one hour and twenty-five minutes, processing around 550,000 tokens, and returned a design featuring moving steering, doors, a folding top, and an engine cover. Regardless of its utility to you, this demonstrates that long, unattended work sessions are becoming commonplace. The issue isn't about working hard; our main argument is that the real scarcity lies in the judgment surrounding that work.

02Delegation is the real interface

When an agent operates around the clock, your focus shifts from doing work to defining tasks. The important audit before activating one of these agents includes:

• —Delegate tasks with clear, checkable outcomes first. For example, “Monitor these feeds and flag changes against this definition” can operate independently overnight, while “Improve our positioning” cannot.
• —Clearly write down what the agent may decide autonomously. Set thresholds, caps, response templates, and escalation paths. If you can’t articulate the boundaries, the agent won’t be able to follow them.
• —Determine what necessitates human intervention. Everything else — the checks it performs and the retries it makes — falls under configuration. The list of reasons for waking you up is essentially policy.

This does not involve prompt design, which is why it contrasts with the idea of "a glorified Codex session": the product is not the session itself; it’s the underlying framework that surrounds it.

03Who watches the watcher?

A practitioner pointed out this week that as agents gain autonomy, governance, observability, approvals, and guardrails cease to be mere enterprise paperwork and evolve into the product itself. An agent equipped with its own computer, memory, and access to your tools behaves like a new employee, lacking any institutional memory of your rules — except for those you explicitly encode. We publish our delegation guidelines for delivery work, but in this always-on era, they become essential for everyone, not just builders.

So, before you give a Dot access to your tools, answer this crucial question in writing: what decisions is it allowed to make while you sleep? Everything else is just setup.

Sources: The OpenAI Dots launch posts, the amplification from the “hidden video,” the “glorified Codex session” review, the post about governance implications, and the LEGO design that took 1 hour and 25 minutes and used 550,000 tokens — all observed on LinkedIn from September 27 to 30, 2026.