I.
A common picture of where agentic systems might be headed is an always-on personal agent or “Chief of Staff”. We have started the baby steps on this with Claude Duet, ChatGPT Dots, Instinct, Muse, etc; an agent (or an army of them) that knows your goals, has access to all of your tools, and handles things for you. One version of the world might be that once these systems get good enough, durable UIs go away; why click around in an app when you can just ask in chat?
It turns out, many of us already work with an army of general (quite smart) intelligence agents who know our goals and have access to our tools! They’re the people I work with, and, through one lens, I’ve been “running loops” with them for years. So if I want to guess how I’d behave with an AGI chief of staff, a reasonable place to start might be how I behave today with the human intelligences already sitting in those seats.
II.
I have an executive assistant who manages my calendar, and there isn’t anything stopping me from excising the Outlook calendar fully from my life. I could run my whole day via Slack - my assistant could send me my schedule each morning. I could ask for updates (both simple and complex) as the day went on, and I’d quickly get an answer back. I imagine some folks do work this way, but for me this seems silly, and not only because I’d be bashful to send so many dumb questions via Slack! There is plenty of intelligence on the other end of that Slack thread and I still open the calendar, because it’s often much easier to go look at my day, or make a simple change (like a simple cancellation). Of course, there are requests I make via Slack (often much more complex calendaring requirements), but by volume, I expect I use the calendar UI at least as often as Slack.
I’d expect the same with an agent doing the scheduling: I hand off the work and still want to look at the result myself. In ClassPass, MCP, and the absence of UX moats I found that composing tools at the LLM layer beat the app for a surprising number of interactions, while photos, long-tail search results, and filters were better in a graphical experience “for now”. The EA analogy makes me think I was too quick with that hedge for at least part of the list, since a smarter model does nothing to change how much faster I can read a grid than a paragraph. I’d guess the part that does move over to the agent is the input side; all of the clicking and typing it takes to get something onto the calendar in the first place.
III.
The prediction I find more interesting is about building software. Once agents are good enough, why does anyone need to be in the loop at all after the goal has been set?
Well, again, I already run /loop "Accomplish these goals" with human intelligences. We have teams of engineers who I trust to write great code and make the right architectural decisions. Instead, I check in on a longer cadence (generally every six weeks). In that conversation I’m trying to understand what they’ve been up to, what they think they should work on next, and where they’re having trouble so that I can help troubleshoot. I can also bring in other context or connections that they might not have at their altitude, which might affect what we build or how we build it. Every so often there is an architectural decision important enough that I ask to be inserted more directly into the process, knowing full well that this will slow the team down. My CEO does the same thing with me, on an even longer loop.
One way of thinking about an organization is as a set of nested loops with different periods. An engineer’s loop runs in hours or days, their manager’s might run every day or every sprint, and mine with their teams is every six weeks. Each one of these loops is already operated by a general intelligence - and the reasons that a manager would check in on a team of agents, or I would check in on an agentic manager are the same reasons that we do so today with smart and motivated individuals - because a goal is set outside of the loop that executes it and has a tendency to drift, and because whether the work is right mostly depends on context about a more complex system than just the work itself (this is what the mind meld and agreeing on what good looks like are for, and I’d want both even with people I trust completely to do the job).1
So I would expect the same with an army of AGI agents. Swapping an engineer’s inner loop for an agent moves work faster, and swapping a team of engineers for a team of agents means that six weeks of work might happen in an afternoon2 - but doing so doesn’t necessarily remove the outer loops. An agent can be as capable as my best engineer and still not know that the priorities changed, or that a different project has started that might require a change in approach.
IV.
Agents aren’t people, of course. An agent is cheaper to interrupt than a team, and (as far as I know) it doesn’t get demoralized when I opine on its architecture. All of that lowers the cost of looping in, which means you might expect to insert higher-level loops more often into work done by agents than those done by humans.