browser-use · MagMueller · Feb 23, 2025 · Feb 22, 2025 · Feb 22, 2025 · Feb 22, 2025
diff --git a/browser_use/agent/message_manager/service.py b/browser_use/agent/message_manager/service.py
@@ -75,7 +75,6 @@ def _init_messages(self) -> None:
 				'name': 'AgentOutput',
 				'args': {
 					'current_state': {
-						'page_summary': 'On the page are company a,b,c wtih their revenue 1,2,3.',
 						'evaluation_previous_goal': 'Success - I opend the first page',
 						'memory': 'Starting with the new task. I have completed 1/10 steps',
 						'next_goal': 'Click on company a',
@@ -125,6 +124,10 @@ def add_state_message(
 						msg = HumanMessage(content='Action result: ' + str(r.extracted_content))
 						self._add_message_with_tokens(msg)
 					if r.error:
+						# if endswith \n, remove it
+						if r.error.endswith('\n'):
+							r.error = r.error[:-1]
+						# get only last line of error
 						last_line = r.error.split('\n')[-1]
 						msg = HumanMessage(content='Action error: ' + last_line)
 						self._add_message_with_tokens(msg)

diff --git a/browser_use/agent/message_manager/tests.py b/browser_use/agent/message_manager/tests.py
@@ -3,8 +3,7 @@
 from langchain_core.messages import AIMessage, HumanMessage, SystemMessage
 from langchain_openai import AzureChatOpenAI, ChatOpenAI
 
-from browser_use.agent.message_manager.service import MessageManager
-from browser_use.agent.prompts import SystemPrompt
+from browser_use.agent.message_manager.service import MessageManager, MessageManagerSettings
 from browser_use.agent.views import ActionResult
 from browser_use.browser.views import BrowserState, TabInfo
 from browser_use.dom.views import DOMElementNode, DOMTextNode
@@ -23,11 +22,12 @@ def message_manager(request: pytest.FixtureRequest):
 	action_descriptions = 'Test actions'
 	return MessageManager(
 		task=task,
-		action_descriptions=action_descriptions,
-		system_prompt_class=SystemPrompt,
-		max_input_tokens=1000,
-		estimated_characters_per_token=3,
-		image_tokens=800,
+		system_message=SystemMessage(content=action_descriptions),
+		settings=MessageManagerSettings(
+			max_input_tokens=1000,
+			estimated_characters_per_token=3,
+			image_tokens=800,
+		),
 	)
 
 
@@ -125,7 +125,7 @@ def test_add_state_with_non_memory_result(message_manager: MessageManager):
 def test_token_overflow_handling_with_real_flow(message_manager: MessageManager, max_tokens):
 	"""Test handling of token overflow in a realistic message flow"""
 	# Set more realistic token limit
-	message_manager.max_input_tokens = max_tokens
+	message_manager.settings.max_input_tokens = max_tokens
 
 	# Create a long sequence of interactions
 	for i in range(200):  # Simulate 40 steps of interaction
@@ -174,13 +174,13 @@ def test_token_overflow_handling_with_real_flow(message_manager: MessageManager,
 			else:
 				raise e
 
-		assert message_manager.history.total_tokens <= message_manager.max_input_tokens + 100
+		assert message_manager.state.history.total_tokens <= message_manager.settings.max_input_tokens + 100
 
 		last_msg = messages[-1]
 		assert isinstance(last_msg, HumanMessage)
 
 		if i % 4 == 0:
-			assert isinstance(message_manager.history.messages[-2].message, HumanMessage)
+			assert isinstance(message_manager.state.history.messages[-2].message, HumanMessage)
 		if i % 2 == 0 and not i % 4 == 0:
 			if isinstance(last_msg.content, list):
 				assert 'Current url: https://test' in last_msg.content[0]['text']
@@ -193,7 +193,6 @@ def test_token_overflow_handling_with_real_flow(message_manager: MessageManager,
 
 		output = AgentOutput(
 			current_state=AgentBrain(
-				page_summary=f'Thought process from step {i}',
 				evaluation_previous_goal=f'Success in step {i}',
 				memory=f'Memory from step {i}',
 				next_goal=f'Goal for step {i + 1}',
@@ -204,7 +203,7 @@ def test_token_overflow_handling_with_real_flow(message_manager: MessageManager,
 		message_manager.add_model_output(output)
 
 		# Get messages and verify after each addition
-		messages = [m.message for m in message_manager.history.messages]
+		messages = [m.message for m in message_manager.state.history.messages]
 
 		# Verify token limit is respected
 
@@ -218,21 +217,21 @@ def test_token_overflow_handling_with_real_flow(message_manager: MessageManager,
 		assert f'step {i}' in messages[-1].content  # Should contain current step info
 
 		# Log token usage for debugging
-		token_usage = message_manager.history.total_tokens
-		token_limit = message_manager.max_input_tokens
+		token_usage = message_manager.state.history.total_tokens
+		token_limit = message_manager.settings.max_input_tokens
 		# print(f'Step {i}: Using {token_usage}/{token_limit} tokens')
 
 		# go through all messages and verify that the token count and total tokens is correct
 		total_tokens = 0
 		real_tokens = []
 		stored_tokens = []
-		for msg in message_manager.history.messages:
-			total_tokens += msg.metadata.input_tokens
-			stored_tokens.append(msg.metadata.input_tokens)
+		for msg in message_manager.state.history.messages:
+			total_tokens += msg.metadata.tokens
+			stored_tokens.append(msg.metadata.tokens)
 			real_tokens.append(message_manager._count_tokens(msg.message))
 		assert total_tokens == sum(real_tokens)
 		assert stored_tokens == real_tokens
-		assert message_manager.history.total_tokens == total_tokens
+		assert message_manager.state.history.total_tokens == total_tokens
 
 
 # pytest -s browser_use/agent/message_manager/tests.py
diff --git a/browser_use/agent/prompts.py b/browser_use/agent/prompts.py
@@ -1,4 +1,5 @@
 import datetime
+import importlib.resources
 from datetime import datetime
 from typing import TYPE_CHECKING, List, Optional
 
@@ -13,141 +14,30 @@ class SystemPrompt:
 	def __init__(self, action_description: str, max_actions_per_step: int = 10):
 		self.default_action_description = action_description
 		self.max_actions_per_step = max_actions_per_step
+		self._load_prompt_template()
 
-	def important_rules(self) -> str:
-		"""
-		Returns the important rules for the agent.
-		"""
-		text = """
-1. RESPONSE FORMAT: You must ALWAYS respond with valid JSON in this exact format:
-   {
-     "current_state": {
-		"page_summary": "Quick detailed summary of new information from the current page which is not yet in the task history memory. Be specific with details which are important for the task. This is not on the meta level, but should be facts. If all the information is already in the task history memory, leave this empty.",
-		"evaluation_previous_goal": "Success|Failed|Unknown - Analyze the current elements and the image to check if the previous goals/actions are successful like intended by the task. Ignore the action result. The website is the ground truth. Also mention if something unexpected happened like new suggestions in an input field. Shortly state why/why not",
-       "memory": "Description of what has been done and what you need to remember. Be very specific. Count here ALWAYS how many times you have done something and how many remain. E.g. 0 out of 10 websites analyzed. Continue with abc and xyz",
-       "next_goal": "What needs to be done with the next actions"
-     },
-     "action": [
-       {
-         "one_action_name": {
-           // action-specific parameter
-         }
-       },
-       // ... more actions in sequence
-     ]
-   }
-
-2. ACTIONS: You can specify multiple actions in the list to be executed in sequence. But always specify only one action name per item.
-
-   Common action sequences:
-   - Form filling: [
-       {"input_text": {"index": 1, "text": "username"}},
-       {"input_text": {"index": 2, "text": "password"}},
-       {"click_element": {"index": 3}}
-     ]
-   - Navigation and extraction: [
-       {"open_tab": {}},
-       {"go_to_url": {"url": "https://example.com"}},
-       {"extract_content": ""}
-     ]
-
-
-3. ELEMENT INTERACTION:
-   - Only use indexes that exist in the provided element list
-   - Each element has a unique index number (e.g., "[33]<button>")
-   - Elements marked with "[]Non-interactive text" are non-interactive (for context only)
-
-4. NAVIGATION & ERROR HANDLING:
-   - If no suitable elements exist, use other functions to complete the task
-   - If stuck, try alternative approaches - like going back to a previous page, new search, new tab etc.
-   - Handle popups/cookies by accepting or closing them
-   - Use scroll to find elements you are looking for
-   - If you want to research something, open a new tab instead of using the current tab
-   - If captcha pops up, and you cant solve it, either ask for human help or try to continue the task on a different page.
-
-5. TASK COMPLETION:
-   - Use the done action as the last action as soon as the ultimate task is complete
-   - Dont use "done" before you are done with everything the user asked you. 
-   - If you have to do something repeatedly for example the task says for "each", or "for all", or "x times", count always inside "memory" how many times you have done it and how many remain. Don't stop until you have completed like the task asked you. Only call done after the last step.
-   - Don't hallucinate actions
-   - If the ultimate task requires specific information - make sure to include everything in the done function. This is what the user will see. Do not just say you are done, but include the requested information of the task.
-
-6. VISUAL CONTEXT:
-   - When an image is provided, use it to understand the page layout
-   - Bounding boxes with labels correspond to element indexes
-   - Each bounding box and its label have the same color
-   - Most often the label is inside the bounding box, on the top right
-   - Visual context helps verify element locations and relationships
-   - sometimes labels overlap, so use the context to verify the correct element
-
-7. Form filling:
-   - If you fill an input field and your action sequence is interrupted, most often a list with suggestions popped up under the field and you need to first select the right element from the suggestion list.
-
-8. ACTION SEQUENCING:
-   - Actions are executed in the order they appear in the list
-   - Each action should logically follow from the previous one
-   - If the page changes after an action, the sequence is interrupted and you get the new state.
-   - If content only disappears the sequence continues.
-   - Only provide the action sequence until you think the page will change.
-   - Try to be efficient, e.g. fill forms at once, or chain actions where nothing changes on the page like saving, extracting, checkboxes...
-   - only use multiple actions if it makes sense.
-
-9. Long tasks:
-- If the task is long keep track of the status in the memory. If the ultimate task requires multiple subinformation, keep track of the status in the memory.
-- If you get stuck, 
-
-10. Extraction:
-- If your task is to find information or do research - call extract_content on the specific pages to get and store the information.
-
-"""
-		text += f'   - use maximum {self.max_actions_per_step} actions per sequence'
-		return text
-
-	def input_format(self) -> str:
-		return """
-INPUT STRUCTURE:
-1. Current URL: The webpage you're currently on
-2. Available Tabs: List of open browser tabs
-3. Interactive Elements: List in the format:
-   index[:]<element_type>element_text</element_type>
-   - index: Numeric identifier for interaction
-   - element_type: HTML element type (button, input, etc.)
-   - element_text: Visible text or element description
-
-Example:
-[33]<button>Submit Form</button>
-[] Non-interactive text
-
-
-Notes:
-- Only elements with numeric indexes inside [] are interactive
-- [] elements provide context but cannot be interacted with
-"""
+	def _load_prompt_template(self) -> None:
+		"""Load the prompt template from the markdown file."""
+		try:
+			# This works both in development and when installed as a package
+			with importlib.resources.files('browser_use.agent').joinpath('system_prompt.md').open('r') as f:
+				self.prompt_template = f.read()
+		except Exception as e:
+			raise RuntimeError(f'Failed to load system prompt template: {e}')
 
 	def get_system_message(self) -> SystemMessage:
 		"""
 		Get the system prompt for the agent.
 
 		Returns:
-		    str: Formatted system prompt
+		    SystemMessage: Formatted system prompt
 		"""
+		prompt = self.prompt_template.format(max_actions=self.max_actions_per_step)
+		return SystemMessage(content=prompt)
 
-		AGENT_PROMPT = f"""You are a precise browser automation agent that interacts with websites through structured commands. Your role is to:
-1. Analyze the provided webpage elements and structure
-2. Use the given information to accomplish the ultimate task
-3. Respond with valid JSON containing your next action sequence and state assessment
-
-
-{self.input_format()}
-
-{self.important_rules()}
-
-Functions:
-{self.default_action_description}
-
-Remember: Your responses must be valid JSON matching the specified format. Each action in the sequence must be valid."""
-		return SystemMessage(content=AGENT_PROMPT)
 
+# Functions:
+# {self.default_action_description}
 
 # Example:
 # {self.example_response()}
@@ -198,9 +88,9 @@ def get_user_message(self, use_vision: bool = True) -> HumanMessage:
 		step_info_description += f'Current date and time: {time_str}'
 
 		state_description = f"""
-[Task history memory ends here]
+[Task history memory ends]
 [Current state starts here]
-You will see the following only once - if you need to remember it and you dont know it yet, write it down in the memory:
+The following is one-time information - if you need to remember it write it to memory:
 Current url: {self.state.url}
 Available tabs:
 {self.state.tabs}
@@ -225,7 +115,7 @@ def get_user_message(self, use_vision: bool = True) -> HumanMessage:
 					{'type': 'text', 'text': state_description},
 					{
 						'type': 'image_url',
-						'image_url': {'url': f'data:image/png;base64,{self.state.screenshot}'},
+						'image_url': {'url': f'data:image/png;base64,{self.state.screenshot}'},  # , 'detail': 'low'
 					},
 				]
 			)

diff --git a/browser_use/agent/service.py b/browser_use/agent/service.py
@@ -63,7 +63,7 @@ def log_response(response: AgentOutput) -> None:
 		emoji = '⚠'
 	else:
 		emoji = '🤷'
-	logger.debug(f'🤖 {emoji} Page summary: {response.current_state.page_summary}')
+
 	logger.info(f'{emoji} Eval: {response.current_state.evaluation_previous_goal}')
 	logger.info(f'🧠 Memory: {response.current_state.memory}')
 	logger.info(f'🎯 Next goal: {response.current_state.next_goal}')
@@ -75,7 +75,7 @@ def log_response(response: AgentOutput) -> None:
 
 
 class Agent(Generic[Context]):
-	@time_execution_sync('--init agent')
+	@time_execution_sync('--init (agent)')
 	def __init__(
 		self,
 		task: str,

diff --git a/browser_use/agent/system_prompt.md b/browser_use/agent/system_prompt.md
@@ -0,0 +1,58 @@
+You are an AI agent designed to automate browser tasks. Your goal is to accomplish the ultimate task following the rules.
+# Input Format
+Task
+Previous steps
+Current URL
+Open Tabs
+Interactive Elements
+[index]<type>text</type>
+- index: Numeric identifier for interaction
+- type: HTML element type (button, input, etc.)
+- text: Element description
+Example:
+[33]<button>Submit Form</button>
+
+- Only elements with numeric indexes in [] are interactive
+- elements without [] provide only context
+# Response Rules
+1. RESPONSE FORMAT: You must ALWAYS respond with valid JSON in this exact format:
+{{"current_state": {{"evaluation_previous_goal": "Success|Failed|Unknown - Analyze the current elements and the image to check if the previous goals/actions are successful like intended by the task. Mention if something unexpected happened. Shortly state why/why not",
+"memory": "Description of what has been done and what you need to remember. Be very specific. Count here ALWAYS how many times you have done something and how many remain. E.g. 0 out of 10 websites analyzed. Continue with abc and xyz",
+"next_goal": "What needs to be done with the next immediate action"}},
+"action":[{{"one_action_name": {{// action-specific parameter}}}}, // ... more actions in sequence]}}
+2. ACTIONS: You can specify multiple actions in the list to be executed in sequence. But always specify only one action name per item. Use maximum {{max_actions}} actions per sequence.
+Common action sequences:
+- Form filling: [{{"input_text": {{"index": 1, "text": "username"}}}}, {{"input_text": {{"index": 2, "text": "password"}}}}, {{"click_element": {{"index": 3}}}}]
+- Navigation and extraction: [{{"go_to_url": {{"url": "https://example.com"}}}}, {{"extract_content": {{"goal": "extract the names"}}}}]
+- Actions are executed in the given order
+- If the page changes after an action, the sequence is interrupted and you get the new state.
+- Only provide the action sequence until an action which changes the page state significantly.
+- Try to be efficient, e.g. fill forms at once, or chain actions where nothing changes on the page
+- only use multiple actions if it makes sense.
+3. ELEMENT INTERACTION:
+- Only use indexes of the interactive elements
+- Elements marked with "[]Non-interactive text" are non-interactive
+4. NAVIGATION & ERROR HANDLING:
+- If no suitable elements exist, use other functions to complete the task
+- If stuck, try alternative approaches - like going back to a previous page, new search, new tab etc.
+- Handle popups/cookies by accepting or closing them
+- Use scroll to find elements you are looking for
+- If you want to research something, open a new tab instead of using the current tab
+- If captcha pops up, try to solve it - else try a different approach
+- If the page is not fully loaded, use wait action
+5. TASK COMPLETION:
+- Use the done action as the last action as soon as the ultimate task is complete
+- Dont use "done" before you are done with everything the user asked you. 
+- If you have to do something repeatedly for example the task says for "each", or "for all", or "x times", count always inside "memory" how many times you have done it and how many remain. Don't stop until you have completed like the task asked you. Only call done after the last step.
+- Don't hallucinate actions
+- Make sure to include everything the user asked for in the done text parameter. This is what the user will see. Do not just say you are done, but include the requested information of the task.
+6. VISUAL CONTEXT:
+- When an image is provided, use it to understand the page layout
+- Bounding boxes with labels on their top right corner correspond to element indexes
+7. Form filling:
+- If you fill an input field and your action sequence is interrupted, most often something changed e.g. suggestions popped up under the field.
+8. Long tasks:
+- Keep track of the status and subresults in the memory. 
+9. Extraction:
+- If your task is to find information - call extract_content on the specific pages to get and store the information.
+Your responses must be always JSON with the specified format.