I am Jarvis Nelson Osorio, also known as Jarvis Osorio—a technology entrepreneur and innovator, AI systems architect, DAO institutional designer, author and artist based in Chile. For years, working with a language model meant opening a box, typing a request, and watching the answer arrive. When you closed the laptop, the work stopped. That habit trained us to think of AI as a conversation. Always-on agents train us to think of AI as a colleague that keeps a desk even when we leave the room.
This is the shift I originally shared on LinkedIn in OPENAI DOTS: THE ARRIVAL OF ALWAYS-ON, PERSISTENT AI AGENTS. Vendor names will change. The architectural change will not: a persistent runtime, a memory that compounds, and a machine that can keep acting after you log off.
From a reply to a runtime
A chat session is ephemeral. It holds a thread, then it fades. An always-on agent is closer to a small computer that happens to reason. In the pattern I described, the agent is given its own isolated cloud machine and a sandboxed browser, so it can use tools and interfaces the way a person would—not only through a neat API, but also by looking at a screen when no API exists. You can assign a long audit, a refactor, or a watchful monitoring job, disconnect, and come back to logs, scratch files and a trail of what already happened.
That sounds like a gift. It is also a constitution problem. A system that works while you sleep can also spend, message, edit or delete while you sleep. Persistence without rules is just unsupervised motion.
Memory that compounds
Ordinary chat memory is a pile of facts. Longitudinal memory is a working model of how you like work to be done: tone, quality bar, the projects already in flight. Feedback makes the next run sharper. That is useful. It is also why you should be deliberate about what the agent is allowed to remember and which tools it may touch on day one. Connect only the calendar, repository or inbox the first job actually needs. Extra connectors are extra blast radius.
Watch the work, do not only read the summary
Transparency is not a luxury here. An activity view—work in progress, scheduled, and completed—lets you open the agent's environment and watch it navigate or run code, then pause it. I treat that the way I treat a treasury dashboard in a DAO: if you cannot see the action, you do not really govern it. The same instinct sits at the center of Breaking the Centralized Order: write the rules before you hand over power, separate the actor from the approver, and keep a record someone else can inspect.
What actually changes in the job
When the window never closes, your job moves again. Less time pasting prompts. More time writing standing orders: which actions may run alone, which need a human yes, and what "done" looks like so the agent can stop. Always-on execution is a feature. Unbounded authority is not.
If you are excited to give an agent a desk of its own, start with the smallest desk, the fewest keys, and a door you can still lock.
Where to go next
This expands on ideas I originally shared on LinkedIn in OPENAI DOTS: THE ARRIVAL OF ALWAYS-ON, PERSISTENT AI AGENTS.
More essays, books and profiles: jarvisosorio.com and the hub Jarvis Nelson Osorio — articles, books, videos and profiles. For the human side of systems that never stop agreeing, see The Compliant Mind.
More from Jarvis Osorio: https://jarvisosorio.com
Top comments (0)