Agently Docs

EN

中文

EN

中文

Appearance

Browse Tool ​

Applies to: v4.0.8.2

Browse is now the only built-in browsing tool. Playwright and PyAutoGUI are no longer exposed as standalone tools; they are internal backends of Browse.

1. Fallback decision chain ​

How to read this diagram ​

  • fallback_order decides ordering, but each stage is still gated by enable_* flags and environment availability.
  • “Failure” is not only exceptions. Content that is too short also counts as a failed backend attempt.

2. Why there is one Browse tool instead of many browsing tools ​

This is an architectural simplification, not a capability reduction:

  • business logic only says “I need reliable page content”
  • backend selection is centralized in Browse
  • fallback strategy, content quality checks, and dynamic rendering behavior all stay in one tool

That is why projects like Daily News Collector no longer need application-level branching between Playwright and plain HTML fetching.

3. Key parameters ​

Default backend order:

python
("pyautogui", "playwright", "bs4")

Common setup:

python
from agently.builtins.tools import Browse

browse = Browse(
    enable_pyautogui=False,
    enable_playwright=True,
    enable_bs4=True,
    response_mode="markdown",
    min_content_length=80,
)

Important parameters:

  • fallback_order
  • enable_pyautogui
  • enable_playwright
  • enable_bs4
  • response_mode
  • min_content_length
  • max_content_length
  • playwright_headless
  • playwright_timeout
  • playwright_include_links
  • pyautogui_open_mode

4. Content extraction and failure semantics ​

Design rationale ​

The core of Browse is not “opening a page”. It is “returning text evidence that is long enough and content-like enough to be useful”. That is what makes it better than a naive httpx + BeautifulSoup.get_text() approach for docs sites and news sites.

Under the bs4 path, it prefers:

  • main/article/.vp-doc/[role=main] style content roots
  • filtering nav/aside/header/footer
  • filtering sidebar/toc/navbar-style noise selectors

5. Return value and failure contract ​

Public API:

python
content = await browse.browse(url)

On success:

  • extracted content string

On failure:

  • Can not browse '...'. Fallback failed: ...

So browse() still returns a string, not a structured trace object. Wrap it yourself if you need strict error handling.

6. Why Daily News Collector enables Playwright by default ​

Its SETTINGS.yaml uses:

yaml
BROWSE:
  enable_playwright: true

because many news sites:

  • render useful content only after browser execution
  • rely on dynamic loading
  • produce content that is too thin under plain HTML fetch

With min_content_length, this makes the fallback model especially suitable for news collection.

7. Recommended usage ​

python
from agently import Agently
from agently.builtins.tools import Search, Browse

agent = Agently.create_agent()
search = Search(region="us-en")
browse = Browse(enable_playwright=True, min_content_length=80)

agent.use_tools([search.search, search.search_news, browse.browse])

Best practices:

  • search first, browse second
  • check Browse output length and error strings
  • cite evidence or URLs in the final answer