'Continue? Y/N' — A 60-Second Game Just Exposed AI Agents' Biggest UX Problem

A strange little game made the rounds on developer Twitter last week. It’s called Continue? Y/N. Playtime: about 60 seconds. Calling it a game is generous — it’s closer to performance art. All you do is click “Continue? [Y/N]” prompts that won’t stop appearing. So why did this nothing of a game blow up among engineers? Because everyone who uses AI coding agents took one look and said: yeah, that’s my life.

One click takes a second. Multiply that by 200.

If you’ve used Claude Code, Cursor, or any agentic coding tool, the scene writes itself. Read this file? Allow? Run this bash command? Allow? Create this directory? Allow?

The first ten minutes, you read each prompt carefully. By minute eleven, your hand is on autopilot. Y, Y, Y, Y. Eventually you either smash “Allow All” or — worse — you keep clicking Y on autopilot until you wave through something genuinely dangerous. The safety mechanism becomes the security hole.

The developer turned this into a one-minute interactive bit. Simple Y/N choices at first. Then the prompts speed up. The options get vaguer. By the end, half the screen is consent dialogs and you have no idea what you just agreed to. You just keep pressing Y. Security researchers have a name for this: permission fatigue.

Mobile OSes already solved this problem a decade ago

If this feels familiar, it should. Early-2010s mobile apps did exactly the same thing. Remember when a flashlight app demanded contacts, location, microphone, and camera all at once? Users tapped “Allow” and rivers of data flowed out the back.

Both iOS and Android fixed it with runtime, granular permissions. Stop asking at install time. Ask the moment the app actually needs the capability. Add nuance — “Only this time,” “While using the app,” “Allow with limited access.” It took roughly a decade to get there.

AI agents are currently stuck in 2010. Either you confirm every micro-action, or you hand over the keys upfront. There is no middle.

The deeper problem: agents do things you didn’t predict

Here’s where the mobile analogy breaks down in the agent’s favor — and against it. A mobile app’s capabilities are known in advance. Instagram is never going to suddenly open your banking app.

An AI agent decides what to do on the fly. You say “fix this bug,” and the agent reads ten files, installs a package, runs the test suite, and commits to git. You’re being asked to evaluate, in real time, whether each step is actually necessary. In 60 seconds. Two hundred times.

Cognitive scientists call this decision fatigue. Human judgment isn’t infinite — repeated trivial choices degrade the quality of the important ones. There’s a famous study showing judges grant parole more often right after lunch, because their decision-making capacity recovers with rest. AI agents are essentially asking us to be the pre-lunch judge, all day, every day.

What the fix probably looks like

A few directions are emerging. First, risk-tiered permissions: auto-allow read-only operations, batch-confirm writes, always confirm anything that touches the shell or the network. Some tools — Claude Code’s --allowedTools, Cursor’s auto-mode — are crawling in this direction.

Second, sandboxing. Run the agent inside an isolated environment (a container, a worktree, a VM) and let it do whatever it wants. Permission prompts become unnecessary because the blast radius is contained.

Third, and most ambitious: intent-based approval. Approve a high-level goal once — “ship this PR” — and let the agent operate freely within a pre-declared scope. Only prompt when the agent wants to step outside the lines. This is, more or less, the road mobile already walked.

What a 60-second game is really asking

The reason this thing resonated isn’t that it’s funny — though it is. It held up a mirror to the people building AI tools. The safety mechanisms we built in the name of user control may actually be a way of offloading responsibility. “You clicked Y” is a great alibi. It’s not great product design.

Try counting today. How many times did you press Y on an AI tool? How many of those did you actually read? That gap is the empty seat where the next AI UX breakthrough belongs.