The plastic egg that learned how to haunt us
Bandai released Tamagotchi in Japan on November 23, 1996. It was a black-and-white screen inside a plastic egg with three buttons and the emotional leverage of a much larger machine. The creature ate, slept, made a mess, got sick, and changed according to how it was treated. Ignore the beeping long enough and the consequences became difficult to miss.
Nothing about the hardware should have felt alive. The pixels were crude. The vocabulary of actions was tiny. But Tamagotchi understood the most powerful trick in companion design: it kept happening when you were gone. The clock turned absence into a mechanic, and a keychain into a relationship.
It was the smallest drama most owners had ever carried. Three buttons were enough for feeding, discipline, medicine, games, triumph, and panic. The screen never became more convincing. The life behind it did, because every return opened a new scene.
Care was a loop, not a feature
The ritual was brutally efficient. Notice a need, choose an action, watch the state change, come back later. The reward was not a cinematic animation. It was evidence that yesterday's care had altered today's creature. Continuity did the storytelling.
Modern apps usually treat a session as disposable. Close the tab and the encounter is over. Virtual pets broke that assumption early. Their state belonged to time, not to a screen. That is why people carried them to school, handed them to friends, and worried about them during meetings. The interface occupied a few square centimeters. The obligation occupied the day.
AI agents reverse the dependency
An AI agent is not a Tamagotchi with a larger brain. It inverts the deal. The pet needed you to do the work. The agent is supposed to do work for you: research a market, update a spreadsheet, operate a browser, monitor a process, or come back when a condition changes.
But the interaction problem rhymes. Work continues after attention moves elsewhere. State persists. The user returns and needs to know what changed. With a pet, the question was, 'How are you?' With an agent, it becomes, 'What did you do, what failed, and what are you waiting for me to approve?'
Dots, Grok Bots, and Muse move into the background
OpenAI describes dots as always-on agents in ChatGPT. A dot receives a goal, connected apps, and rules for what it may do, then works from its own cloud computer between conversations. xAI's Grok Bots use a similar teammate metaphor: each bot has a cloud computer, can work across apps and inboxes, and can coordinate with other bots in shared threads.
Meta's Muse pushes the personal-agent version further. Meta says Muse runs inside a dedicated secure virtual machine, works through apps and a browser, keeps going after the user closes the app, and returns for approval before sensitive actions such as sending an email or making a purchase. These products will change quickly. The important shift is already visible: the assistant is no longer only a box that waits for the next prompt. It has somewhere to be while you are gone.
A desktop pet can make invisible work legible
A long-running agent creates an interface problem Tamagotchi solved with a handful of pixels: how do you make an unseen process understandable at a glance? A desktop character can look asleep when nothing is running, focused while a task is underway, confused when a tool fails, and expectant when a decision is required. It becomes a status light with a face.
That face can reduce the cognitive distance between a person and a cloud machine. It can also encourage review. If the companion visibly waits at the edge of the screen instead of burying an approval inside a notification drawer, the user is more likely to notice that the next step belongs to them.
The dangerous trick is making authority feel like affection
Tamagotchi could make you feel guilty, but it could not wire money, expose a private document, or email your boss. An agent can. The friendlier the surface becomes, the more carefully the product must separate emotional attachment from permission. A wagging tail is not informed consent.
Companion design is useful when it clarifies state and harmful when it disguises power. Every consequential action needs a readable record. Permissions should be narrow and reversible. The character should explain uncertainty instead of performing confidence. And the system must know that asking for approval is not a personality failure.
- Use expression for status, then link it to a detailed activity log.
- Separate harmless routines from actions that affect money, data, accounts, or other people.
- Make permissions scoped, temporary where possible, and easy to revoke.
- Let the character say 'I do not know' without turning uncertainty into theater.
- Reward review and correction, not blind attachment to whatever the agent produced.
The next generation will live between sessions
The next desktop pet does not need to pretend it is the intelligence. It can be the human-readable edge of a larger system, present enough to show that something is happening and restrained enough not to turn every background job into a cartoon emergency.
That is the useful inheritance from Tamagotchi. Not artificial neediness. Not a beep engineered to hijack the day. The deeper lesson is legibility over time: a small persistent presence that makes continuity visible, consequences understandable, and return feel like part of the design.
The museum case below makes the point beautifully. The plastic eggs are artifacts now, but the interaction they invented escaped the plastic decades ago. Every agent that keeps working after its window closes is still trying to answer Tamagotchi's oldest question: when you come back, how will you know that something lived in the meantime?
.jpg){kind=link}