AI Browser Agent Troubleshooting: 10 Things That Go Wrong

TL;DR: Most AI Browser Agent failures come from five places: the model setup, a vague goal, the limits you set, the site itself, and sign-ins. The agent tells you which one it hit in the result message. This guide lists the messages you’ll see, what caused them and what to change.

Cereal’s AI Browser Agent drives your own Chrome toward a goal you write in plain language. It reads the page, decides what to click, and either answers you (Once) or keeps checking until a condition holds (Repeat). Most of the time that works. When it doesn’t, the cause is usually one of the ten below.

Each section starts with the message the agent shows, so you can search this page for the text you got.

1. The model isn’t reachable or rejects the request

You see: Couldn’t reach your AI provider, rejected the request, or returned an error.

The agent has no model of its own. It uses the one you set up in Settings → AI, and if that call fails, the run stops before the agent does anything on the page.

Rejected the request almost always means a wrong or expired API key. Copy it again from your provider, with no spaces around it, and click Test connection.
Couldn’t reach means the provider didn’t answer: you’re offline, the base URL is wrong, or a local model server isn’t running. For Repeat tasks the agent retries on the next check before giving up.
Returned an error shows the provider’s own message. Rate limits on free tiers show up here. Wait, pick a model with higher limits, or move to a paid tier. Our free Gemini key guide covers the free-tier limits.

2. The goal is too vague

You see: The model stopped without an answer, a give-up reason from the agent, or an answer to a slightly different question than the one you meant.

The goal is the whole instruction. “Check the pliers” leaves the agent guessing: which pliers, and check what? Good goals name the thing, the value you want and what counts as done:

Weak: check the hammer price
Better: tell me the price of the cheapest claw hammer in the Hand Tools category

If the information really isn’t on the page, the agent fails on purpose rather than making something up. That’s a feature. Check that the start URL leads to the page that has the answer.

3. The run mode doesn’t fit the goal

You see: Repeat mode needs a check, or Couldn’t find something on the page to watch.

Once and Repeat want different kinds of goals:

Once suits a question: tell me whether Long Nose Pliers is in stock.
Repeat suits a condition that can become true: notify me when Long Nose Pliers is in stock.

On its first run, a Repeat task learns one element on the page to watch, like a stock label or a price. If the goal doesn’t point at anything that can change, there’s nothing to watch. Rewrite it as “notify me when …” and name what should change.

4. It hit a step or dollar limit

You see: Stopped after N steps, the Max steps per run limit, Stopped at $2.00, the Max $ per run limit, or This task has spent $5.00, its Max $ per task limit.

Every model call is one step, and every step costs money at your provider. The defaults stop a run at $2 and a task at $5, so a confused agent can’t run up a bill.

If a run keeps stopping at a limit, raising the limit is rarely the fix. Usually the goal sends the agent through too many pages. Start it closer to the answer: a category or product page instead of the home page.

Watch out for models Cereal doesn’t have a price for, such as local models or new releases. Their cost shows as unknown and counts as $0, so Max steps per run is the only limit that applies to them.

5. It keeps repeating the same action

You see: The agent got stuck repeating the same action on a page.

When the agent clicks the same button with the same input three times on one page without progress, it gets a warning. If it keeps going, the run stops. This usually happens on a button that doesn’t react: a disabled “Add to cart” until you choose a size, a filter that reloads the same list, or a “Load more” that loads nothing.

Tell the goal about the hidden step, such as pick size 42 first, or start on a URL that already has the filter applied.

6. The site sends it to another domain

You see: is not an allowed domain.

The agent only goes to the start URL’s own domain, plus the domains you allow. That keeps it from wandering across the web. It also means a login page on a separate domain, such as login.example-id.com, or a payment provider is off limits until you allow it.

Add the domain under Allowed domains in the task. The same rule protects your sign-in secrets: the agent won’t type them on a page outside the allowed domains.

You see: The goal needs a secret ‘…’, but Sign-in secrets has no row with that label, or Signed out and cleared the learned steps.

Sign-in secrets let the agent type your email and password without the AI ever seeing them. Two things trip people up:

The label doesn’t match. If the goal says log in with my email and password, the task needs rows labeled email and password. Labels are lowercase with _ for spaces.
It got stuck on the sign-in page. A wrong password, a 2FA prompt or a “verify it’s you” screen keeps the agent on the login page. When that happens it signs out and forgets its learned steps, so the next start is clean. Fix the secret, or sign in yourself once in the task’s Chrome window, which keeps its own profile between runs.

8. The site changed

You see: The saved steps stopped matching … The agent takes over, or The page changed, so I re-learned it for $0.04. Still watching.

This one isn’t really a failure. A Repeat task learns the steps on its first run and replays them later without calling the model, which is why routine checks are free. When the site’s layout changes, replay stops where it no longer fits and the model takes over from there. You’ll see a small cost for that check, then free checks again.

If a site redesigns often, expect that small cost now and then. If the learned steps don’t fit at all anymore, the agent forgets them and learns from scratch.

9. The site can’t be reached or blocks bots

You see: Couldn’t open … Check the Start URL, Couldn’t load …, or The monitor stopped: 3 checks in a row couldn’t reach …

A typo in the start URL, a site that’s down, or your own connection all show up here. A Repeat task shrugs off one or two failed checks. After three in a row it stops, so it doesn’t keep checking a dead site.

Some sites block automated browsers outright or show a CAPTCHA. The agent doesn’t solve CAPTCHAs. On sites with heavy bot protection, a ready-made Cereal script built for that site is often the better tool.

10. Part of the page is invisible to it

You see: an answer that ignores something you can clearly see, or a fail saying the information isn’t there.

The agent reads the page as structured text, not as a picture. Content inside an embedded frame from another site, such as a third-party chat widget, a review widget or a payment form, shows up only as a placeholder. Content drawn on a canvas, or text that exists only inside an image, isn’t readable either.

If the value you want lives in one of those, the agent can’t get at it. Open the embedded page directly as the start URL if it has its own address.

Bonus: it waited for you

If a run seems to hang before buying, paying, deleting or sending something, check for a Confirm dialog. The agent asks before every irreversible step. If you decline, you’ll see Stopped: you declined to … and nothing was done.

Still stuck?

Each run’s log in Cereal shows every step the agent took and what it saw. Your secrets are removed from it. Start there. If the log doesn’t explain it, bring the result message and your goal to our Discord and we’ll help you sort it out.

New to the agent? Set it up with a free Gemini API key first, then download Cereal and sign in. The AI Browser Agent is added to your account automatically and installs right after you sign in.