Skip to content

Open-source Python library for AI browser agents, plus Browser Use Cloud: stealth browsers and a hosted agent API.

update agent success with judge verdict

Saurav Pandadae6dc6d0e7e
diff --git a/browser_use/agent/service.py b/browser_use/agent/service.py
index 398096fa1..1da5e252f 100644
--- a/browser_use/agent/service.py
+++ b/browser_use/agent/service.py
@@ -1322,7 +1322,7 @@ class Agent(Generic[Context, AgentStructuredOutput]):
return None
async def _judge_and_log(self) -> None:
- """Run judge evaluation and log the verdict"""
+ """Run judge evaluation, override agent success with judge verdict, and log."""
judgement = await self._judge_trace()
# Attach judgement to last action result
@@ -1333,8 +1333,11 @@ class Agent(Generic[Context, AgentStructuredOutput]):
# Get self-reported success
self_reported_success = last_result.success
- # Log the verdict based on self-reported success and judge verdict
if judgement:
+ # Override agent's success with judge verdict
+ if judgement.verdict is not None:
+ last_result.success = judgement.verdict
+
# If both self-reported and judge agree on success, don't log
if self_reported_success is True and judgement.verdict is True:
return
@@ -1342,7 +1345,7 @@ class Agent(Generic[Context, AgentStructuredOutput]):
judge_log = '\n'
# If agent reported success but judge thinks it failed, show warning
if self_reported_success is True and judgement.verdict is False:
- judge_log += '⚠️ \033[33mAgent reported success but judge thinks task failed\033[0m\n'
+ judge_log += '⚠️ \033[33mAgent reported success but judge thinks task failed — overriding to failure\033[0m\n'
# Otherwise, show full judge result
verdict_color = '\033[32m' if judgement.verdict else '\033[31m'