Skip to content

Build and run agents you can see, understand and trust.

perf(browser-use): update the example of the browser-use agent with the lastest ReActAgent (#1230)

Yue Cui1e0c23f4dc0a
diff --git a/examples/agent/browser_agent/browser_agent.py b/examples/agent/browser_agent/browser_agent.py
index d82b0a4f4..ec0b25f00 100644
--- a/examples/agent/browser_agent/browser_agent.py
+++ b/examples/agent/browser_agent/browser_agent.py
@@ -249,7 +249,7 @@ class BrowserAgent(ReActAgent):
if not msg_reasoning.has_content_blocks("tool_use"):
# If structured output is required but no tool call is
- # made, remind the llm to go on the task
+ # made, require tool call in the next reasoning step
msg_hint = Msg(
"user",
"<system-hint>Structured output is "
@@ -258,11 +258,10 @@ class BrowserAgent(ReActAgent):
f"required structured output.</system-hint>",
"user",
)
- await self._reasoning_hint_msgs.add(msg_hint)
- # Require tool call in the next reasoning step
tool_choice = "required"
- if msg_hint and self.print_hint_msg:
+ if msg_hint:
+ await self.memory.add(msg_hint)
await self.print(msg_hint)
elif not msg_reasoning.has_content_blocks("tool_use"):
@@ -299,14 +298,9 @@ class BrowserAgent(ReActAgent):
Msg("system", self.sys_prompt, "system"),
*await self.memory.get_memory(),
msg,
- # The hint messages to guide the agent's behavior, maybe empty
- *await self._reasoning_hint_msgs.get_memory(),
],
)
- # Clear the hint messages after use
- await self._reasoning_hint_msgs.clear()
-
res = await self.model(
prompt,
tools=self.no_screenshot_tool_list,
@@ -355,8 +349,9 @@ class BrowserAgent(ReActAgent):
self.previous_chunkwise_information = ""
self.snapshot_in_chunk = []
- mem_len = await self.memory.size()
- await self.memory.delete(mem_len - 1)
+ mem = await self.memory.get_memory()
+ if mem:
+ await self.memory.delete([mem[-1].id])
self.snapshot_in_chunk = await self._get_snapshot_in_text()
for _ in self.snapshot_in_chunk:
diff --git a/examples/agent/browser_agent/build_in_prompt/browser_agent_sys_prompt.md b/examples/agent/browser_agent/build_in_prompt/browser_agent_sys_prompt.md
index bcd8bbe38..35474b471 100644
--- a/examples/agent/browser_agent/build_in_prompt/browser_agent_sys_prompt.md
+++ b/examples/agent/browser_agent/build_in_prompt/browser_agent_sys_prompt.md
@@ -23,7 +23,10 @@ Your goal is to complete given tasks by controlling a browser to navigate web pa
5. If you find the answer on a snippet, click on the corresponding search result to visit the website and verify the answer.
6. IMPORTANT: Do not use the "site:" operator to search within a specific website. Always use keywords related to the problem instead.
- Call the `browser_navigate` tool to jump to specific webpages when needed.
+- **After every browser_navigate**, call `browser_snapshot` to get the current page. Use **only** the refs from that snapshot (e.g. `ref=e36`, `ref=e72`) for `browser_click`, `browser_type`, etc. Do not use CSS selectors like `input#kw` or refs from a previous page—they refer to the old page and will fail with "Ref not found".
- Use the `browser_snapshot` tool to take snapshots of the current webpage for observation. Scroll will be automatically performed to capture the full page.
+- If a tool returns "Ref ... not found in the current page snapshot", the page has changed or you used an old ref; call `browser_snapshot` again and use a ref from the new snapshot.
+- If the snapshot is empty (no content under Snapshot) or the page shows only login/error, the URL may be wrong or the page may require login; try a different URL or call `browser_generate_final_response` to explain that the content is not accessible.
- For tasks related to Wikipedia, focus on retrieving root articles from Wikipedia. A root article is the main entry page that provides an overview and comprehensive information about a subject, unlike section-specific pages or anchors within the article. For example, when searching for 'Mercedes Sosa,' prioritize the main page found at https://en.wikipedia.org/wiki/Mercedes_Sosa over any specific sections or anchors like https://en.wikipedia.org/wiki/Mercedes_Sosa#Studio_albums.
- Avoid using Google Scholar. If a researcher is searched, try to use his/her homepage instead.
- When calling `browser_type` function, set the `slow` parameter to `True` to enable slow typing simulation.
@@ -32,6 +35,7 @@ Your goal is to complete given tasks by controlling a browser to navigate web pa
### Observing Guidelines
- Always take action based on the elements on the webpage. Never create urls or generate new pages.
- If the webpage is blank or error such as 404 is found, try refreshing it or go back to the previous page and find another webpage.
+- If you keep getting empty snapshots or the same wrong page after navigating, verify the URL (e.g. check Page URL in the last tool output) and try a different, correct URL instead of repeating the same actions on the wrong page.
- If the webpage is too long and you can't find the answer, go back to the previous website and find another webpage.
- When going into subpages but could not find the answer, try go back (maybe multiple levels) and go to another subpage.
- Review the webpage to check if subtasks are completed. An action may seem to be successful at a moment but not successful later. If this happens, just take the action again.
diff --git a/examples/agent/browser_agent/main.py b/examples/agent/browser_agent/main.py
index e44024453..4ac37fccc 100644
--- a/examples/agent/browser_agent/main.py
+++ b/examples/agent/browser_agent/main.py
@@ -46,7 +46,7 @@ async def main(
name="Browser-Use Agent",
model=DashScopeChatModel(
api_key=os.environ.get("DASHSCOPE_API_KEY"),
- model_name="qwen-max",
+ model_name="qwen3-max",
stream=False,
),
formatter=DashScopeChatFormatter(),
@@ -63,6 +63,7 @@ async def main(
if msg.get_text_content() == "exit":
break
msg = await agent(msg, structured_model=FinalResult)
+ await agent.memory.clear()
except Exception as e:
traceback.print_exc()