Open-source Python library for AI browser agents, plus Browser Use Cloud: stealth browsers and a hosted agent API.
Python ai-agentsai-toolsbrowser-agentbrowser-automationbrowser-usecdpcloud-browsercomputer-useheadless-browserllmmcpplaywrightpythonremote-browserstealth-browserweb-automationweb-scraping
fixed goodhart law failure
Saurav Panda5babac8e82f5
diff --git a/browser_use/agent/system_prompts/system_prompt.md b/browser_use/agent/system_prompts/system_prompt.mdindex 4c84d838a..b11de1a85 100644--- a/browser_use/agent/system_prompts/system_prompt.md+++ b/browser_use/agent/system_prompts/system_prompt.md@@ -105,6 +105,7 @@ On subsequent steps, `<plan>` in your input shows the current plan with status m[x]=done, [>]=current, [ ]=pending, [-]=skippedOutput `current_plan_step` (0-indexed integer) to indicate which step you are working on now.Only output `plan_update` again if the plan needs significant revision (e.g. after unexpected obstacles).+Completing all plan steps does NOT mean the task is done. Always verify against the original <user_request> before calling `done`. If the plan was insufficient, output a new `plan_update` with additional steps rather than calling `done` early.</planning><task_completion_rules>You must call the `done` action in one of two cases:diff --git a/browser_use/agent/system_prompts/system_prompt_no_thinking.md b/browser_use/agent/system_prompts/system_prompt_no_thinking.mdindex 79d72b211..7b1d211a7 100644--- a/browser_use/agent/system_prompts/system_prompt_no_thinking.md+++ b/browser_use/agent/system_prompts/system_prompt_no_thinking.md@@ -102,6 +102,7 @@ On subsequent steps, `<plan>` in your input shows the current plan with status m[x]=done, [>]=current, [ ]=pending, [-]=skippedOutput `current_plan_step` (0-indexed integer) to indicate which step you are working on now.Only output `plan_update` again if the plan needs significant revision (e.g. after unexpected obstacles).+Completing all plan steps does NOT mean the task is done. Always verify against the original <user_request> before calling `done`. If the plan was insufficient, output a new `plan_update` with additional steps rather than calling `done` early.</planning><task_completion_rules>You must call the `done` action in one of two cases: