Multitasking in the Age of AI Agents: Attention and Recovery

Posted on In

I have been working on several research and software projects at the same time. With AI coding agents, there is now another layer of multitasking: a handful of sessions can be running in parallel, each occasionally asking me to review a change, answer a question, or decide what to do next.

This creates an interesting tension. More work can happen in parallel, but my attention still has to move from one context to another. I wanted to understand what psychology says about that experience—and, more importantly, how to organize a day so that the extra parallelism is useful without leaving me mentally exhausted.

My guiding idea is simple: let projects and agents run concurrently, while giving my own attention room to focus and recover.

Two Timescales of Multitasking

I find it useful to distinguish two kinds of multitasking.

Long-horizon work Short-horizon interactions
Timescale Days, weeks, months Seconds, minutes, hours
Typical work Research projects, courses, software development Reading agent updates, reviewing diffs, answering questions
Main challenge Keeping several projects moving over time Switching attention and reconstructing context
Useful habit Weekly planning and daily focus Asynchronous execution and deliberate review windows

The number of active projects, the number of running agent sessions, and the number of times I switch attention are three different things. Keeping many projects alive does not require visiting all of them every hour. Likewise, agents can run concurrently without requiring continuous human supervision.

What Task Switching Costs

Switch Cost

Cognitive psychology has a well-established concept called switch cost. In experiments, people tend to respond more slowly—and sometimes make more errors—immediately after switching from one task to another. Even when a switch is predictable and people have time to prepare, some cost remains 1.

Real-world tasks are more complicated than the simple tasks used in laboratory experiments. When I reopen a coding session, I may need to recall the goal of the change, the current architecture, the tests that failed, and the reasoning behind the previous decision. Going from a paper to a different software project can involve a substantial reconstruction of context.

Attention Residue

Another useful concept is attention residue. After switching from unfinished work A to work B, part of my attention may remain occupied by A. Sophie Leroy's research describes how this lingering attention can impair performance on the new task 2.

This feels especially relevant when several agent sessions contain open questions. I can move my eyes to a different window immediately; mentally leaving the previous problem may take longer.

Does Multitasking Harm the Brain?

The clearest effects of task switching are immediate: slower responses, occasional mistakes, and the effort of getting back into a task. What concerns me more over time is the combination of high workload, frequent interruptions, a feeling of being constantly on call, and insufficient recovery. The World Health Organization identifies excessive workloads, long hours, and limited control over work among the psychosocial risks to mental health.

Why Watching Agent Sessions All Day Can Be Exhausting

A day with several agents can create three overlapping demands:

  1. Context reconstruction. Each session has its own goal, files, constraints, recent decisions, and unresolved problems.
  2. Attention residue. After reviewing one session, I may continue thinking about it while trying to work on another.
  3. Continuous vigilance. If I expect the next agent message to require an immediate answer, even otherwise quiet time can feel like waiting for an interruption.

The key distinction is between an agent doing work and an agent needing me. An agent can be busy for twenty minutes while I read a paper. When it produces a result, that result can often wait for the next review period.

Long-Horizon Work: Several Projects per Week, Fewer Deep Switches per Day

I would like several projects to progress over the course of a week. That does not mean I need to make meaningful progress on every project every day.

An observational study of GitHub developers, The Sky Is Not the Limit: Multitasking Across GitHub Projects, is interesting here. It distinguishes participation in multiple projects from switching between many projects in a short period 3. That distinction helps me think about how to divide my week.

A reasonable rhythm for me is:

  • Weekly: keep a short overview of active projects, their next milestones, and what is waiting on someone or something else.
  • Daily: pick a primary project or question that deserves an uninterrupted block of thought.
  • Between blocks: handle lighter maintenance work and review results from other projects.
  • At transitions: leave a short note about where to resume.

Projects can remain active even during days when they receive little of my attention. This also leaves room for useful cross-project connections without turning the entire day into a series of shallow visits.

The Three-Line Ready-to-Resume Note

Leroy and Glomb studied a simple intervention called a ready-to-resume plan: before leaving an interrupted task, spend a moment planning how to return to it. This helped reduce attention residue 4.

For example, before leaving a software task:

Current state: The implementation is done; I am reviewing the diff.

Open question: One rollback path still needs checking.

Next action: Inspect the failing test, then review the exceptional branch.

This takes roughly a minute. When I return, I can resume from a clear entry point instead of rereading the entire history.

Supervising Multiple Agents

Treat Agent Activity as Asynchronous Work

I want my agents to keep working even when I am doing something else. Their outputs can form a queue of items that need attention.

Agent state What I can do
Running Let it work; keep my attention on the current task.
Completed Add its result to the next review batch.
Waiting for input Queue the question; prioritize it if it blocks important progress.
Failed or blocked Assess the impact and decide when to intervene.
High-risk action requiring approval Review explicitly before allowing the action.

An urgent failure may deserve immediate attention. For routine updates, I can choose when to step in instead of following every status change.

Three Complementary Modes

I find it helpful to think in terms of three modes:

Focus. Reading a paper, reasoning about a design, or writing something difficult. This is where a sustained train of thought matters most. A possible starting block is 45–90 minutes.

Supervise. Reviewing agent output, answering questions, making decisions, and dispatching the next round of tasks. A possible starting block is 15–30 minutes.

Recover. Getting up, moving around, drinking water, or simply being away from the screen. Even 5–10 minutes can offer a real break.

I can adjust these blocks to the work, my energy level, and the day's schedule.

Notifications: Planned Checks and Genuine Exceptions

There is a difference between these two workflows:

Reactive: an agent replies → I open it → return to the paper → another agent replies → I open that one.

Deliberate: agents produce updates → routine updates accumulate → I review them together at a convenient point.

Research on batching smartphone notifications found improvements in some measures of well-being and perceived control compared with receiving notifications continuously 5. I would like to try the same idea with routine agent updates.

In either case, I would keep an exception for genuinely time-sensitive or safety-critical matters. Everything else can usually wait until I have reached a natural stopping point.

What a Day Might Look Like

I do not need to schedule my attention with the precision of an operating-system scheduler. A few recognizable phases are enough.

Phase My focus What the agents can do
Start of day Choose the main objective and launch independent tasks. Begin autonomous work.
Deep-work block Read, write, think; leave routine notifications alone. Continue running in the background.
Review window Read results, inspect diffs, respond, and delegate next steps. Receive feedback and continue.
Short break Step away from the screen. Continue or wait; no supervision needed.
End of day Leave resumption notes and close the workday. Non-urgent work can wait until tomorrow.

The important change is who controls the rhythm. I would like to check agents because I have decided it is a good moment to do so, not because every new message resets my priorities.

Rest Is More Than Switching to Another Task

When reading a paper makes me tired, checking an agent's code might feel like a break because it is different work. But it still requires attention, decisions, and screen time.

A genuine pause might mean walking around, stretching, drinking water, looking outside, or sitting quietly for a few minutes. The point is to let the mind disengage from active problem solving.

A systematic review and meta-analysis of micro-breaks found reductions in fatigue and increases in vigor. Effects on performance in complex cognitive tasks were less certain 6. That is a useful expectation to set: a break can be worthwhile because it helps me feel and function better, even if it does not immediately make me solve a problem faster.

Staying Alert and Mentally Healthy

Work organization matters, but so do the basics of physical recovery.

Sleep

The CDC recommends at least seven hours of sleep per night for adults aged 18–60. Sleep supports alertness, learning, memory, and emotional regulation.

One tempting habit with agents is to keep checking them late at night because they are still running. I can let them run while giving myself a definite stopping point. The next review can happen after sleep.

Daylight and Movement

Getting outside during the day, moving away from the desk, and maintaining a regular schedule help support a healthier daily rhythm. For physical activity, the WHO recommends 150–300 minutes of moderate-intensity aerobic activity per week (or the vigorous-intensity equivalent), plus muscle-strengthening activity on at least two days 7.

This is particularly worth remembering during weeks spent mostly reading, coding, and reviewing agent output.

Caffeine

Caffeine can improve short-term alertness, but it is not a substitute for sleep. Late caffeine can interfere with the following night's rest, so I want to be mindful of timing and individual sensitivity.

A Real End to the Workday

It is easier to recover when the workday has an ending. I can leave open questions, pending reviews, and next steps in a trusted place, then spend time on activities unrelated to the current projects.

The long-term goal is to be alert and engaged when working, and able to relax when resting.

Signs That My Work Rhythm Needs Adjusting

Counting every context switch is probably less useful than observing how I feel and function.

Some signs worth noticing are:

  • I open a session and forget what I was trying to accomplish.
  • Sustained reading or thinking feels increasingly difficult.
  • I repeatedly check whether an agent has replied, even during rest.
  • I feel mentally tired early in the day and do not recover after breaks.
  • My sleep worsens, or I stay preoccupied with work after stopping.
  • I become persistently irritable, anxious, or less interested in things I usually enjoy.

If difficulties persist for weeks or significantly interfere with everyday life, talking with a healthcare or mental-health professional would be sensible.

A Small Seven-Day Experiment

I would start by changing only three habits for a week:

  1. Keep one uninterrupted focus block each day. I can run agents in parallel without actively checking them.
  2. Batch routine agent reviews. Before changing tasks, leave a three-line ready-to-resume note.
  3. Protect real breaks and sleep. Include time away from the screen and a consistent end-of-day boundary.

At the end of each day, I can give three simple subjective scores from 1 to 5: focus, fatigue, and sleep quality. After seven days, I can see what helps and adjust my rhythm accordingly.

Three Things I Want to Remember

Several projects can progress concurrently; deep attention benefits from continuity.

AI agents can operate asynchronously; I can choose when to supervise them.

Sustainable productivity includes sleep, movement, breaks, and genuine recovery.

The best outcome of having more capable agents would be more freedom to think carefully—and more freedom to stop thinking about work when the day is over.

References

Additional background: APA on switching costs · WHO on mental health at work · CDC on sleep.

  1. Rogers, R. D., & Monsell, S. (1995). Costs of a Predictable Switch Between Simple Cognitive Tasks. Journal of Experimental Psychology: General.↩︎

  2. Leroy, S. (2009). Why Is It So Hard to Do My Work? The Challenge of Attention Residue When Switching Between Work Tasks. Organizational Behavior and Human Decision Processes.↩︎

  3. Vasilescu, B., et al. (2016). The Sky Is Not the Limit: Multitasking Across GitHub Projects. ICSE.↩︎

  4. Leroy, S., & Glomb, T. M. (2018). Tasks Interrupted: How Anticipating Time Pressure on Resumption of an Interrupted Task Causes Attention Residue and Low Performance on Interrupting Tasks and How a “Ready-to-Resume” Plan Mitigates the Effects. Organization Science.↩︎

  5. Fitz, N., et al. (2019). Batching Smartphone Notifications Can Improve Well-being. Computers in Human Behavior.↩︎

  6. Albulescu, P., et al. (2022). “Give Me a Break!” A Systematic Review and Meta-analysis on the Efficacy of Micro-breaks for Increasing Well-being and Performance. PLOS ONE.↩︎

  7. World Health Organization (2020). Guidelines on Physical Activity and Sedentary Behaviour.↩︎