← Blog

I tested it for two days: 20 seconds after Claude in Chrome goes idle, its background pages freeze

Why AI agents working in Chrome's background tabs slow down, get frozen, and can even get a support chat disconnected.

Chat ended: no check-in from this page for 83 s.

That was the last line to pop up on a mock support chat page I built.

The "support agent" had just said "Found the double charge, reply YES and I'll refund it today." I asked Claude to reply YES for me in Chrome. It went off for a little over a minute and a half, and by the time it came back, the chat had been closed.

I built that page to reproduce something that happened to me this September.

On September 19 I gave Claude three jobs at once in the browser through the Claude in Chrome extension. One was checking our browser's popup-detection feature, one was finding music for a video, and one was getting me a refund from a seller on Amazon Australia, because a small keyboard I'd bought there had lost a key after less than two months.

I really liked that keyboard. It's tiny and easy to take with me, so the missing key annoyed me more than it should have. The seller first said no, then stopped answering altogether, and I had to go to Amazon support.

Once all three jobs were under way, I left the browser in the background and went back to the terminal to keep working on an architecture design with Claude Code.

When I switched back a while later, only one of the three jobs was moving. The other two had stopped as if they'd fallen asleep.

I was puzzled and just stared at the screen for a bit. Agents are supposed to get these jobs done on their own. If an agent can't work on several things in parallel, that's a huge flaw!

Most people's first guess would be a slow network, or the model having a bad day.

But if the network or the model were the problem, all three jobs should have slowed down together, and one of them was still moving.

I've worked on browsers for many years, and the first thing that came to mind was: Chrome is throttling the background tabs.

I tested it by hand the next day and confirmed that was the problem. Recently I reproduced it properly, over two days of testing. When I wrote it up, I found the mechanism is fairly complicated and hard to explain cleanly. Simplified: Chrome has three rules for pages nobody is looking at, which I'll call throttling, freezing, and no rendering (the page isn't drawn).

And now the story has a twist. On November 12 the Claude in Chrome extension, which has 22 million users on the Chrome Web Store, got an update. Its listing now says: "Scheduled tasks: set recurring browser work that runs on your schedule." In other words, work that happens while you're not at the screen. Yeah, right.

1. Throttling: a 90-second wait, still not done after six minutes

When an agent works in the background, does Chrome treat background tabs differently?

Whichever tab Claude in Chrome is operating on, it tells Chrome "someone is looking at this tab", so the tab being operated on is exempt from throttling and freezing. Every other tab, as far as Chrome is concerned, is a background page nobody is looking at.

I made a very ordinary download page that says "Your download will start in 90 seconds". The button only becomes clickable when the countdown reaches 0, and it's written exactly like the countdowns on flash-sale and payment pages.

Then I gave Claude two identical tasks, both to wait for the countdown and then download, in two tabs, A and B, in the same window. I minimized the window and had Claude work in tab A often and leave tab B alone.

Tab A was fine and could download at 90 seconds. Tab B slowed down. In two runs, when I checked it after the page had been open for over six minutes, its 90-second countdown had only got through a little over a minute.

Switching back after the pages have been open for over six minutes. Left, Chrome: the countdown in the background tab, which should have finished in 90 seconds, still isn't done. Right, Embiody: it finished at 90 seconds.

Why was B that slow?

If you've written front-end code, think of it this way: the countdown is a recursive setTimeout chain, where each second's callback schedules the wait for the next second. Chrome's first rule, throttling, targets exactly this kind of timer. About a minute after nobody is looking at a page, those timers can only fire once a minute.

It's like a digital watch that should tick every second but breaks and ticks once a minute instead, and as long as nobody comes back to look at it, it keeps going that slowly.

That's what happened to the two stalled jobs back in September.

I opened a countdown page by hand and left it in the background for 8 minutes, and the countdown fell six and a half minutes behind. Anyone running an agent in Chrome runs into this rule.

But slow is at least something you can wait out. The next one felt more serious to me.

2. Freezing: 20 seconds idle, and the page stops completely

The Claude in Chrome extension puts the tabs it opens into one group, a run of colored tabs on the tab strip, and that group can collapse into a small block.

New tabs it opens go behind the tab you're looking at by default. Claude reads them and clicks in them, but never brings them to the front, so these tabs have never been shown on screen.

I had it open a new test tab and left it untouched. After only 20 seconds the page froze completely. It didn't even move once a minute, and it only woke up when Claude did something in it.

I tried it four times, and every time it froze at about 20 seconds. The window wasn't minimized or covered; the tab was just in the background. Why would the page stop?

I watched the tab strip for a while and finally caught an animation: about 20 seconds after Claude's last command finished, its tab group collapsed by itself.

About 20 seconds after Claude goes idle, its tab group collapses on its own, and the test tab in it is frozen right after.

Once a tab group collapses, Chrome applies its second rule, freezing, to the tabs in the group that have never been shown on screen.

What does frozen mean?

Think of Megatron frozen in the Arctic ice. He isn't dead and his memory is all still there, but not a single line of code runs until someone wakes him up.

The extension's collapsing and Chrome's freezing each make sense on their own. The extension doesn't want its pile of tabs hogging your tab strip, so it tucks the group away when it goes idle, and only leaves it open when one of the group's tabs is showing at the front of the window. Chrome figures you don't need anything that's been tucked away for now, so it freezes it to save memory.

Put the two sensible things together, though, and the moment Claude takes a breather, every page it opened in the background stops, and Claude has no idea.

You might think: why not just have the extension stop collapsing?

Put yourself in the shoes of Claude in Chrome's product manager, and it's a dilemma. If it doesn't collapse, its pile of tabs sits across your tab strip and gets in your way. If it does, Chrome freezes them. Collapsing is something it can change; freezing is Chrome's call. It can only pick one of the two.

The good news is there's an exception: the tab Claude touched last can sometimes escape being frozen.

We went through the Claude in Chrome extension's code. Counting from when Claude stops operating a tab, at around the 20-second mark two things happen: the extension withdraws the "someone is looking" signal, and the tab group collapses.

In other words, only if Claude keeps operating every tab it opened can it keep them from freezing. If the agent decides it only needs to come back to a tab "in a while", the group collapses and the tab is frozen right away.

But that's not how agents work. They stop to think for a bit, wait for you to finish running a command in the terminal, wait for a sluggish page, and any of those can take tens of seconds. Once it pauses for more than 20 seconds, every tab except the last one it touched is frozen.

Keeping tabs awake across several tasks takes luck. And nobody dares hand anything important to a tab that stays awake by luck.

3. Support never got the YES

If a countdown freezes, fine, it just runs late, and picks up again whenever the agent remembers to come back to that tab.

But that refund with Amazon support in September dragged on through several attempts before it got sorted, because of the freeze timing out the chat, on top of the usual back-and-forth explaining.

The "Chat ended" at the top of this post is the test I built around that problem.

Pages like live support chats need you to stay present. Every few seconds they quietly tell the server "I'm still here", and if the other end hears nothing for too long, it assumes you've left and closes the session.

I built a support page like that. It checks in every 10 seconds, and the server ends the chat if it gets no check-in for 45 seconds. A support chat's check-in usually schedules the next one only after the server replies, rather than chaining one timer onto the last, so throttling doesn't touch it.

Claude stopped for 20 seconds, the tab group collapsed, and the support page was frozen along with it.

By the time Claude came back, the page had been silent for over 80 seconds. The server had closed the chat back at 45 seconds, and what popped up was the line at the top.

The send button was greyed out. The YES could never be sent, and the refund they had promised for today went nowhere.

When I saw that red text, my first thought was: good thing this is a test page. My second thought was: good thing I now use BrowserDo, the MCP I built, with the Embiody agent browser, which doesn't have this problem.

I ran the same task in Embiody, with Claude operating through BrowserDo instead, away for the same minute and a half and more. The support page kept checking in every 10 seconds and the YES went through.

The agent comes back to the support page after a little over a minute and a half away. Left, Chrome: the chat has ended and YES can't be sent. Right, Embiody: still connected, YES sent.

4. No rendering: the page is open, but the number it reads is wrong

If that were all, it would already be a big problem. But there's a third rule, and it's sneakier.

Plenty of sites like to show numbers with animation. A corporate homepage might say "98,765 companies served", counting up from 0 when you open it and settling on the final figure two or three seconds later.

I made a page like that, counting up to 98,765 in 3 seconds, had Claude open it in a new background tab, and come back 10 seconds later to read the number. That's still under 20 seconds, so the page wasn't frozen; it just looked slow:

Setup Number the agent read
Chrome + Claude in Chrome, three runs 73,795, 71,215, 69,619
Embiody + BrowserDo, two runs 98,765, 98,765
A stats page counts from 0 to 98,765 in 3 seconds. Left, Chrome: the background tab stays at 0 and only starts counting when the agent comes to read it, so the agent reads 69,619. Right, Embiody: the agent reads 98,765.

It wasn't frozen, so why did Chrome never read 98,765, not even once?

Animation on a web page needs the browser to draw it frame by frame. The third rule, no rendering, means Chrome barely draws new frames for a tab that isn't on screen. So this page sat open for 10 seconds with the number stuck at 0, and the animation never even started.

When Claude came back to read it, it told Chrome "someone is looking", and only then did the animation start. But it takes 3 seconds to finish, Claude didn't wait, and all three times it read a number about seventy percent of the way there.

The worst part is that the agent won't tell you the number might be wrong.

From where it sits, the page opened, the number was read, all fine, and it hands you the wrong number as fact.

5. If you want to keep running agents in Chrome, do this for now

These three rules won't change any time soon. Until then, here's what you can do. I tested the first two; the last two are worked out from how the rules work.

To avoid freezing, give Claude's tab group its own window: press and hold the colored group name on the tab strip, drag it out of the window and let go, and the whole group becomes a new window. The tab at the front of that window is one of its own, so the extension won't collapse the group. I tested it: the support page stayed up for 11 minutes, checked in about every 10 seconds, and never dropped once.

You can also put the page that really matters at the front of the window. Once it's been shown, it won't be frozen; even if the window is minimized, it's only throttled.

To avoid reading wrong numbers, when reading an animated number or a page that just opened, have the agent wait a few seconds and read it again, and only trust it if both reads match.

To avoid throttling, you can add the launch flags --disable-background-timer-throttling, --disable-renderer-backgrounding and --disable-backgrounding-occluded-windows after the "Target" in your Chrome shortcut. But I checked against the Chromium source, and no launch flag turns off the no-rendering rule.

To see whether your own browser does this, open https://embiody.com/lab/background-tab/ , switch to another tab for two minutes, come back, and look at "Behind by". That's how many seconds the countdown fell behind while you were away.

6. AI agents need a browser of their own

All of these tricks are, in the end, gaps found in Chrome's behavior. They're stopgaps. Chrome was designed for people, and that's where its architecture starts from.

Chrome's rules were built around "one person at the screen, looking at one tab at a time", and for all these years they've served people well.

But now there's often nobody at the screen. Behind the tabs is an agent looking after several pages at once, waiting on downloads, holding support chats, reading data for you.

In the AI era, people and agents sharing the browser is where things are clearly heading.

After that hand test in September, we decided to add a feature to Embiody, the browser we build for agents, called "AI tabs", so agents don't have to look for gaps. Any tab an agent opens in it is an AI tab. Whether or not the agent is operating it right now, and whether or not the window is minimized, the countdowns in it keep running, and pages the agent reads in the background keep being drawn.

Here are the comparisons from the three sections above side by side:

Test Chrome + Claude in Chrome Embiody + BrowserDo
Throttling: agent only works in A; with the page open for over six minutes, how far B's 90-second countdown got Two runs: 64 s, 84 s Already done at 90 s
Freezing: agent leaves the support page for a little over a minute and a half, then comes back to reply Chat ended, reply can't be sent Stayed connected, reply sent
No rendering: background stats page counts to 98,765 in 3 s, agent reads it 10 s later Three runs: 73,795, 71,215, 69,619 Both runs: 98,765

Tabs you open yourself aren't covered by this. In the background they're still throttled under the usual power-saving rules, so having an agent installed doesn't make your computer burn more power.

Next time you have an agent do something like this in the background, don't assume it's working. First ask yourself whether that tab is awake right now, or already frozen.

As for the keyboard: on the last attempt the tab didn't freeze, Amazon support agreed to the refund without much trouble, and the money finally came back on the 22nd.

If you don't watch it, it doesn't work. But an agent's work is exactly the work that happens while you aren't watching.