Skip to content

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

force-llm-output-after-failure

Magnus Müllerf474e24b28a5
diff --git a/browser_use/agent/service.py b/browser_use/agent/service.py
index 22d8873e4..d1dc1e5f5 100644
--- a/browser_use/agent/service.py
+++ b/browser_use/agent/service.py
@@ -179,6 +179,7 @@ class Agent(Generic[Context, AgentStructuredOutput]):
directly_open_url: bool = True,
include_recent_events: bool = False,
sample_images: list[ContentPartTextParam | ContentPartImageParam] | None = None,
+ final_response_after_failure: bool = True,
**kwargs,
):
if llm is None:
@@ -262,6 +263,7 @@ class Agent(Generic[Context, AgentStructuredOutput]):
include_tool_call_examples=include_tool_call_examples,
llm_timeout=llm_timeout,
step_timeout=step_timeout,
+ final_response_after_failure=final_response_after_failure,
)
# Token cost service
@@ -687,7 +689,8 @@ class Agent(Generic[Context, AgentStructuredOutput]):
available_file_paths=self.available_file_paths, # Always pass current available_file_paths
)
- await self._handle_final_step(step_info)
+ await self._force_done_after_last_step(step_info)
+ await self._force_done_after_failure()
return browser_state_summary
@observe_debug(ignore_input=True, name='get_next_action')
@@ -838,7 +841,7 @@ class Agent(Generic[Context, AgentStructuredOutput]):
# Increment step counter after step is fully completed
self.state.n_steps += 1
- async def _handle_final_step(self, step_info: AgentStepInfo | None = None) -> None:
+ async def _force_done_after_last_step(self, step_info: AgentStepInfo | None = None) -> None:
"""Handle special processing for the last step"""
if step_info and step_info.is_last_step():
# Add last step warning if needed
@@ -850,6 +853,19 @@ class Agent(Generic[Context, AgentStructuredOutput]):
self._message_manager._add_context_message(UserMessage(content=msg))
self.AgentOutput = self.DoneAgentOutput
+ async def _force_done_after_failure(self) -> None:
+ """Force done after failure"""
+ # Create recovery message
+ if self.state.consecutive_failures >= self.settings.max_failures and self.settings.final_response_after_failure:
+ msg = f'You have failed {self.settings.max_failures} consecutive times. This is your final step to complete the task or provide what you found. '
+ msg += 'Use only the "done" action now. No other actions - so here your action sequence must have length 1.'
+ msg += '\nIf the task could not be completed due to the failures, set success in "done" to false!'
+ msg += '\nInclude everything you found out for the task in the done text.'
+
+ self.logger.debug('Final recovery step setup')
+ self._message_manager._add_context_message(UserMessage(content=msg))
+ self.AgentOutput = self.DoneAgentOutput
+
async def _get_model_output_with_retry(self, input_messages: list[BaseMessage]) -> AgentOutput:
"""Get model output with retry logic for empty actions"""
model_output = await self.get_model_output(input_messages)
@@ -1277,7 +1293,9 @@ class Agent(Generic[Context, AgentStructuredOutput]):
signal_handler.reset()
# Check if we should stop due to too many failures
- if self.state.consecutive_failures >= self.settings.max_failures:
+ if (
+ self.state.consecutive_failures + int(self.settings.final_response_after_failure)
+ ) >= self.settings.max_failures:
self.logger.error(f'❌ Stopping due to {self.settings.max_failures} consecutive failures')
agent_run_error = f'Stopped due to {self.settings.max_failures} consecutive failures'
break
diff --git a/browser_use/agent/views.py b/browser_use/agent/views.py
index 395c95414..fb85e8973 100644
--- a/browser_use/agent/views.py
+++ b/browser_use/agent/views.py
@@ -49,6 +49,7 @@ class AgentSettings(BaseModel):
include_tool_call_examples: bool = False
llm_timeout: int = 60 # Timeout in seconds for LLM calls
step_timeout: int = 180 # Timeout in seconds for each step
+ final_response_after_failure: bool = True # If True, attempt one final recovery call after max_failures
class AgentState(BaseModel):
@@ -64,6 +65,7 @@ class AgentState(BaseModel):
stopped: bool = False
session_initialized: bool = False # Track if session events have been dispatched
follow_up_task: bool = False # Track if the agent is a follow-up task
+ recovery_attempted: bool = False # Track if final recovery has been attempted
message_manager_state: MessageManagerState = Field(default_factory=MessageManagerState)
file_system_state: FileSystemState | None = None
diff --git a/docs/customize/agent/all-parameters.mdx b/docs/customize/agent/all-parameters.mdx
index 5f4469f67..05f6e217e 100644
--- a/docs/customize/agent/all-parameters.mdx
+++ b/docs/customize/agent/all-parameters.mdx
@@ -21,6 +21,7 @@ mode: "wide"
- `initial_actions`: List of actions to run before the main task without LLM. [Example](https://github.com/browser-use/browser-use/blob/main/examples/features/initial_actions.py)
- `max_actions_per_step` (default: `10`): Maximum actions per step, e.g. for form filling the agent can output 10 fields at once. We execute the actions until the page changes.
- `max_failures` (default: `3`): Maximum retries for steps with errors
+- `final_response_after_failure` (default: `True`): If True, attempt to force one final model call with intermediate output after max_failures is reached
- `use_thinking` (default: `True`): Controls whether the agent uses its internal "thinking" field for explicit reasoning steps.
- `flash_mode` (default: `False`): Fast mode that skips evaluation, next goal and thinking and only uses memory. If `flash_mode` is enabled, it overrides `use_thinking` and disables the thinking process entirely. [Example](https://github.com/browser-use/browser-use/blob/main/examples/getting_started/05_fast_agent.py)