Overview

The Agent class is the core component of Browser Use that handles browser automation. Here are the main configuration options you can use when initializing an agent.

Basic Settings

from browser_use import Agent
from langchain_openai import ChatOpenAI

agent = Agent(
    task="Search for latest news about AI",
    llm=ChatOpenAI(model="gpt-4o"),
)

Required Parameters

  • task: The instruction for the agent to execute
  • llm: A LangChain chat model instance. See LangChain Models for supported models.

Agent Behavior

Control how the agent operates:

agent = Agent(
    task="your task",
    llm=llm,
    controller=custom_controller,  # For custom tool calling
    use_vision=True,              # Enable vision capabilities
    save_conversation_path="logs/conversation"  # Save chat logs
)

Behavior Parameters

  • controller: Registry of functions the agent can call. Defaults to base Controller. See Custom Functions for details.
  • use_vision: Enable/disable vision capabilities. Defaults to True.
    • When enabled, the model processes visual information from web pages
    • Disable to reduce costs or use models without vision support
    • For GPT-4o, image processing costs approximately 800-1000 tokens (~$0.002 USD) per image (but this depends on the defined screen size)
  • save_conversation_path: Path to save the complete conversation history. Useful for debugging.
  • override_system_message: Completely replace the default system prompt with a custom one.
  • extend_system_message: Add additional instructions to the default system prompt.

Vision capabilities are recommended for better web interaction understanding, but can be disabled to reduce costs or when using models without vision support.

Reuse Existing Browser Context

By default browser-use launches its own builtin browser using playwright chromium. You can also connect to a remote browser or pass any of the following existing playwright objects to the Agent: page, browser_context, browser, browser_session, or browser_profile.

These all get passed down to create a BrowserSession for the Agent:

agent = Agent(
    task='book a flight to fiji',
    llm=llm,
    browser_profile=browser_profile,  # use this profile to create a BrowserSession
    browser_session=BrowserSession(   # use an existing BrowserSession
      cdp_url=...,                      # remote CDP browser to connect to
      # or
      wss_url=...,                      # remote wss playwright server provider
      # or
      browser_pid=...                   # pid of a locally running browser process to attach to
      # or
      executable_path=...               # provide a custom chrome binary path
      # or
      channel=...                       # specify chrome, chromium, ms-edge, etc.
      # or
      page=page,                        # use an existing playwright Page object
      # or
      browser_context=browser_context,  # use an existing playwright BrowserContext object
      # or
      browser=browser,                  # use an existing playwright Browser object
    ),
)

For example, to connect to an existing browser over CDP you could do:

agent = Agent(
    ...
    browser_session=BrowserSession(cdp_url='http://localhost:9222'),
)

For example, to connect to a local running chrome instance you can do:

agent = Agent(
    ...
    browser_session=BrowserSession(browser_pid=1234),
)

See Connect to your Browser for more info.

You can reuse the same BrowserSession after an agent has completed running. If you do nothing, the browser will be automatically closed on run() completion only if it was launched by us.

Running the Agent

The agent is executed using the async run() method:

  • max_steps (default: 100) Maximum number of steps the agent can take during execution. This prevents infinite loops and helps control execution time.

Agent History

The method returns an AgentHistoryList object containing the complete execution history. This history is invaluable for debugging, analysis, and creating reproducible scripts.

# Example of accessing history
history = await agent.run()

# Access (some) useful information
history.urls()              # List of visited URLs
history.screenshots()       # List of screenshot paths
history.action_names()      # Names of executed actions
history.extracted_content() # Content extracted during execution
history.errors()           # Any errors that occurred
history.model_actions()     # All actions with their parameters

The AgentHistoryList provides many helper methods to analyze the execution:

  • final_result(): Get the final extracted content
  • is_done(): Check if the agent completed successfully
  • has_errors(): Check if any errors occurred
  • model_thoughts(): Get the agent’s reasoning process
  • action_results(): Get results of all actions

For a complete list of helper methods and detailed history analysis capabilities, refer to the AgentHistoryList source code.

Run initial actions without LLM

With this example you can run initial actions without the LLM. Specify the action as a dictionary where the key is the action name and the value is the action parameters. You can find all our actions in the Controller source code.

initial_actions = [
	{'open_tab': {'url': 'https://www.google.com'}},
	{'open_tab': {'url': 'https://en.wikipedia.org/wiki/Randomness'}},
	{'scroll_down': {'amount': 1000}},
]
agent = Agent(
	task='What theories are displayed on the page?',
	initial_actions=initial_actions,
	llm=llm,
)

Run with message context

You can configure the agent and provide a separate message to help the LLM understand the task better.

from langchain_openai import ChatOpenAI

agent = Agent(
    task="your task",
    message_context="Additional information about the task",
    llm = ChatOpenAI(model='gpt-4o')
)

Run with planner model

You can configure the agent to use a separate planner model for high-level task planning:

from langchain_openai import ChatOpenAI

# Initialize models
llm = ChatOpenAI(model='gpt-4o')
planner_llm = ChatOpenAI(model='o3-mini')

agent = Agent(
    task="your task",
    llm=llm,
    planner_llm=planner_llm,           # Separate model for planning
    use_vision_for_planner=False,      # Disable vision for planner
    planner_interval=4                 # Plan every 4 steps
)

Planner Parameters

  • planner_llm: A LangChain chat model instance used for high-level task planning. Can be a smaller/cheaper model than the main LLM.
  • use_vision_for_planner: Enable/disable vision capabilities for the planner model. Defaults to True.
  • planner_interval: Number of steps between planning phases. Defaults to 1.

Using a separate planner model can help:

  • Reduce costs by using a smaller model for high-level planning
  • Improve task decomposition and strategic thinking
  • Better handle complex, multi-step tasks

The planner model is optional. If not specified, the agent will not use the planner model.

Optional Parameters

  • message_context: Additional information about the task to help the LLM understand the task better.
  • initial_actions: List of initial actions to run before the main task.
  • max_actions_per_step: Maximum number of actions to run in a step. Defaults to 10.
  • max_failures: Maximum number of failures before giving up. Defaults to 3.
  • retry_delay: Time to wait between retries in seconds when rate limited. Defaults to 10.
  • generate_gif: Enable/disable GIF generation. Defaults to False. Set to True or a string path to save the GIF.

Memory Management

Browser Use includes a procedural memory system using Mem0 that automatically summarizes the agent’s conversation history at regular intervals to optimize context window usage during long tasks.

from browser_use.agent.memory import MemoryConfig

agent = Agent(
    task="your task",
    llm=llm,
    enable_memory=True,
    memory_config=MemoryConfig( # Ensure llm_instance is passed if not using default LLM config
        llm_instance=llm,      # Important: Pass the agent's LLM instance here
        agent_id="my_custom_agent",
        memory_interval=15
    )
)

Memory Parameters

  • enable_memory: Enable/disable the procedural memory system. Defaults to True.
  • memory_config: A MemoryConfig Pydantic model instance (required if enable_memory is True). Dictionary format is not supported.

Using MemoryConfig

You must configure the memory system using the MemoryConfig Pydantic model for a type-safe approach:

from browser_use.agent.memory import MemoryConfig
from langchain_openai import ChatOpenAI # Assuming llm is an instance of ChatOpenAI

llm_for_agent = ChatOpenAI(model="gpt-4o")

agent = Agent(
    task=task_description,
    llm=llm_for_agent,
    enable_memory=True, # This is True by default
    memory_config=MemoryConfig(
        llm_instance=llm_for_agent, # Pass the LLM instance for Mem0
        agent_id="my_agent",
        memory_interval=15, # Summarize every 15 steps
        embedder_provider="openai",
        embedder_model="text-embedding-3-large",
        embedder_dims=1536,
        # --- Vector Store Customization ---
        vector_store_provider="qdrant", # e.g., Qdrant, Pinecone, Chroma, etc.
        vector_store_collection_name="my_browser_use_memories", # Optional: custom collection name
        vector_store_config_override={ # Provider-specific config
            "host": "localhost",
            "port": 6333
            # Add other Qdrant specific configs here if needed, e.g., api_key for cloud
        }
    )
)

The MemoryConfig model provides these configuration options:

Memory Settings

  • agent_id: Unique identifier for the agent (default: "browser_use_agent"). Essential for persistent memory sessions if using a persistent vector store.
  • memory_interval: Number of steps between memory summarization (default: 10)

LLM Settings (for Mem0’s internal operations)

  • llm_instance: The LangChain BaseChatModel instance that Mem0 will use for its internal summarization and processing. You must pass the same LLM instance used by the main agent, or another compatible one, here.

Embedder Settings

  • embedder_provider: Provider for embeddings ('openai', 'gemini', 'ollama', or 'huggingface')
  • embedder_model: Model name for the embedder
  • embedder_dims: Dimensions for the embeddings

Vector Store Settings

  • vector_store_provider: Choose the vector store backend. Supported options include: 'faiss' (default), 'qdrant', 'pinecone', 'supabase', 'elasticsearch', 'chroma', 'weaviate', 'milvus', 'pgvector', 'upstash_vector', 'vertex_ai_vector_search', 'azure_ai_search', 'lancedb', 'mongodb', 'redis', 'memory' (in-memory, non-persistent).
  • vector_store_collection_name: (Optional) Specify a custom name for the collection or index in your vector store. If not provided, a default name is generated (especially for local stores like FAISS/Chroma) or used by Mem0.
  • vector_store_base_path: Path for local vector stores like FAISS or Chroma (e.g., /tmp/mem0). Default is /tmp/mem0.
  • vector_store_config_override: (Optional) A dictionary to provide or override specific configuration parameters required by Mem0 for the chosen vector_store_provider. This is where you’d put connection details like host, port, api_key, url, environment, etc., for cloud-based or server-based vector stores.

The model automatically sets appropriate defaults based on the LLM being used:

  • For ChatOpenAI: Uses OpenAI’s text-embedding-3-small embeddings
  • For ChatGoogleGenerativeAI: Uses Gemini’s models/text-embedding-004 embeddings
  • For ChatOllama: Uses Ollama’s nomic-embed-text embeddings
  • Default: Uses Hugging Face’s all-MiniLM-L6-v2 embeddings

Important:

  • Always pass a properly constructed MemoryConfig object to the memory_config parameter.
  • Ensure the llm_instance is provided to MemoryConfig so Mem0 can perform its operations.
  • For persistent memory across agent runs or for shared memory, choose a scalable vector store provider (like Qdrant, Pinecone, etc.) and configure it correctly using vector_store_provider and vector_store_config_override. The default ‘faiss’ provider stores data locally in vector_store_base_path.

How Memory Works

When enabled, the agent periodically compresses its conversation history into concise summaries:

  1. Every memory_interval steps, the agent reviews its recent interactions.
  2. It uses Mem0 (configured with your chosen LLM and vector store) to create a procedural memory summary.
  3. The original messages in the agent’s active context are replaced with this summary, reducing token usage.
  4. This process helps maintain important context while freeing up the context window for new information.

Disabling Memory

If you want to disable the memory system (for debugging or for shorter tasks), set enable_memory to False:

agent = Agent(
    task="your task",
    llm=llm,
    enable_memory=False
)

Disabling memory may be useful for debugging or short tasks, but for longer tasks, it can lead to context window overflow as the conversation history grows. The memory system helps maintain performance during extended sessions.