An AI agent can keep working after I close my laptop. I’m still figuring out what that means for my working day.
In a recent LinkedIn post, I mentioned that I had been running experiments through AI agents almost around the clock for at least two weeks. That brought me back to a question I had already encountered during my PhD: the expectation of being “always on.” Back then, social media and constant connectivity were part of that discussion. Now, my own experiments give me another reason to ask where work begins and ends.
Is this healthy? Which hours actually count as work? Does the answer depend on whether an experiment produces something useful?
I don’t have settled answers. But I think these questions deserve as much attention as the tools themselves.
One more check before tomorrow
Consider an ordinary scenario: you give an agent a task near the end of your working day. It’ll take a while. You could leave the result until tomorrow, but there is a reasonable argument for checking it sooner. Perhaps it needs clarification. Perhaps it has finished, and one small follow-up would let it make more progress overnight.
Each intervention might take only a few minutes. Together, they can create a working day with a very long tail.
There’s an obvious attraction here. An experiment can move forward while we do something else. Ideas that would previously have stayed on a backlog become affordable to explore. As someone who enjoys solving technical problems, I find that possibility exciting.
Yet an agent that continues working also gives us a continuing reason to return. Even a successful result can become an invitation to start the next task.
Machine time and human attention
I find it useful to distinguish between three things: the time the machine runs, the time a person actively works, and the attention a task occupies between interactions.
Machine runtime is easy to measure. Active work includes defining the task, supplying context, reviewing output, testing assumptions, and correcting mistakes. The third category is harder to describe: wondering whether a run has finished, remembering that a decision is pending, or mentally preparing the next instruction.
Those are different experiences. Eight hours of agent runtime don’t automatically equal eight hours of human work. Equally, ten minutes of visible interaction may tell us very little about how much of someone’s evening a task has occupied.
I’d like to understand that third category better. If the interface is closed but the person is still anticipating the next result, how much distance from work have they actually gained?
This also complicates how we talk about productivity. If an agent completes a useful task overnight, that is a valuable outcome. That doesn’t mean overnight supervision should become the expectation. And if an experiment fails, the effort spent designing and reviewing it still happened. We already understand this in research and engineering; agent-assisted work should not make us forget it.
As an engineering manager, I think the expectations we establish will matter. “You can leave this running” sounds helpful. “Why didn’t you leave something running?” creates a different relationship with the same capability.
It would be easy for spare machine capacity to become an informal obligation: another experiment could have been running, another implementation could have been attempted, another result could have been ready by morning. There’s always another task. A sustainable workflow needs some way to decide when enough has been delegated for today.
Designing for stepping away
In my previous post about cooperative AI, I argued for systems that support human judgment through useful interaction. That leaves an important design question: when should that interaction happen?
An agent can be good at requesting input and still be inconvenient to work with. If every uncertainty becomes a notification, the human may end up managing a queue of interruptions. Adding an approval button does little to resolve that problem if the person must constantly watch for it.
I’d like to see more explicit agreements about availability. A person should be able to tell a system: continue within this scope, stop at these boundaries, and hold questions until I return. Reaching a boundary should be a normal, visible state. When the agent has nothing left it’s allowed to do, waiting is the correct outcome.
This also connects to my recent project, jev-guard, a safety gate for AI coding agents. It checks the shell commands, file reads, and file writes an agent attempts before they run, and lets them through, blocks them, or requires a human decision. The intention is to let routine work proceed while enforcing boundaries around actions that deserve closer attention.
What interests me here is whether such boundaries make it easier to step away. A system that can stop when necessary offers a different basis for delegation than one that depends on continuous observation. But the balance matters: too many interruptions could recreate the monitoring burden, while too few could create misplaced confidence. I see jev-guard as one practical attempt at that balance. Whether it actually helps people disengage is something worth evaluating.
Coming back matters as much as stepping away. A useful handover would explain what changed, what was checked, where uncertainty remains, and which decision needs attention. That gives the person a reasonable place to resume without reconstructing an entire conversation or execution history.
What I want to try next
Here are some experiments I want to try in my own workflows:
- Can I define a stopping point before starting a long run?
- Which results can wait for a planned review window?
- Can several questions be collected into one reviewable proposal?
- What information would let me leave a task alone with confidence?
- How should a team make clear that a running agent does not imply an available colleague?
These are experiments, not habits I’ve mastered.
There’s also a research opportunity here. Alongside output quality and completion time, I’d like to know how often people check their agents, how much time they spend reviewing results, and whether they feel able to disengage. Do longer autonomous runs give people more uninterrupted time? Under which conditions do they instead encourage more monitoring?
My experience so far has given me questions, not a general answer about what this technology does to our well-being. I remain enthusiastic about what it makes possible. That’s precisely why I want to understand the working patterns we’re building around it.
The possibility of making progress while away from the keyboard is valuable. I want to be able to enjoy that benefit and actually be away.
If you regularly work with AI agents: has delegation made it easier to finish your working day, or given you more reasons to check back?
This post was written with the help of Generative AI.
