Repository navigation
Why browser agents burn tokens: two real stuck runs, and how to catch them #207
tishachawla-jg
started this conversation in
Show and tell
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Browser agents waste most of their money in runs that are stuck, not runs that are hard. Two real runs from our testing:
1. The button that did nothing. Asked for 2023 movie ticket prices on Wolfram|Alpha, the agent clicked Compute on a page that never changed. 25 of its 38 actions were that same click, each reported as
SUCCESS.2. The wall it kept walking into. Asked to look something up on the Cambridge Dictionary, it hit a Cloudflare check on step 4. It waited, switched tabs, tried DuckDuckGo, tried the direct URL, and kept going for more than 30 steps while its own notes said
Failure.What catches this today (released,
pip install agent-tokenops)progress_guardcorrects first, halts only if correcting does not helpprogress_guardwithinput_repeats: 3(opt-in)context_compactiontrims the request before it is sentbudget_paceoutput_runawayWhat is coming (branch, feedback welcome)
feat/browser-agent-governancelets an adapter tell the guard what the agent already knows:progress: "no"from the agent's own step verdict: three in a row count as a stallblocked: "yes"on a bot-check page: an immediate "use another source or finish" messagewrap_up: ask for the best answer so far before haltingAll new options are off by default.
We want your traces
What is the worst loop your agent has gotten into? Reply with the task, what it kept doing, and how many steps it burned. The strange ones shape the next policy. And if you try the branch, tell us what it caught and what it got wrong.
If this is useful, a ⭐ on the repo helps others find it.
All reactions