BrowserGym

BrowserGym copied to clipboard

## Description by Korbit AI ### What change is being made? Add the SafeArena Benchmark to the BrowserGym project by introducing configurations, tasks, metadata, and scripts related to "safearena", updating...

alzambranolu13

the action `scroll_at` was recently added since the scroll function seems to only scroll on the page and not on the active object. If the agent need to scroll inside...

recursix

Doc / tutorials

2

- [ ] tutorial how to implement new tasks / benchmarks - [ ] refer users to AgentLab to evaluate agents

gasse

## Benchmark tutorial This PR serves as a template for benchmark creation. It involves the following steps: - Creating a task object to integrate your tasks to BrowserGym - Registering...

TLSDC

to be compatible with the tool API and MCP, let's make sure the action_set also provide a tool_describe function that returns a json compatible with standard tool API. ```python class...

recursix

Need help for instruction translation

1

Dear BrowserGym team, I really like your work. Is it possible to translate the user's clicking behavior into instructions defined by browsergym? For example, if I click on a button...

XvHaidong

Problem in click action/PDF interaction

1

It seems the click action does not work when agent want to click on View Pdf for archive article. Here is a sample code (parse_content_to_elements, and find_matching_anchor borrowed from [here](https://github.com/TheAgentCompany/TheAgentCompany/blob/main/evaluation/browsing.py))...

DsDastgheib

Right now, backends are decided by a string that is internally matched with a set of accepted strings: ```python benchmark = Benchmark( name="visualwebarena", backends=["visualwebarena"], # ... ) ``` It would...

xhluca

Browsergym-Workarena not supported for local development

2

I've been working with Workarena for a while using the browsergym-workarena Python module. Now, I need to experiment with some of my own proposed task configurations, but to do this,...

Nid989

MiniWoB currently requires manual installation, but could be automated to minimize friction to use.

recursix