ViDA team open-sources BrowserBC, raising agent success on WebArena-Hard to 81% by distilling human browser trajectories

ViDA team open-sources BrowserBC, raising agent success on WebArena-Hard to 81% by distilling human browser trajectories · Digg

Posts from X

Most Activity

VIEWS917RETWEETS13

Vida@vida_agent

We open-sourced BrowserBC: A system that turns human browser trajectories into reusable agent skills. Just one recording is enough to generalize a skill.

🛠️ GitHub: [https://github.com/Einsia/Browser-Journey-Forge]

Here’s how it works. 👇

6h20.5K5649

BOOKMARKS1

Vida@vida_agent

A real booking task. Watch how one human demonstration becomes a reusable skill for another agent.👇

6h8911

LIKES1REPLIES1

Pode vir@thiagoTF

@kimmonismus cheaper models executing distilled skills. so basically agents dont need to think they just need to execute. cool. still want to know who prices the actual demand though

2h451

Vida@vida_agent

BrowserBC doesn’t just improve success rate. It also changes how agents behave.

① Fewer steps per task 31.2 → 22.7 tool calls (−27%)

② Better transfer Strong model → distills skills Small model → executes them well

③ Not a script Skills guide behavior, not execution

Skills reduce exploration. Execution remains the bottleneck. 👇

6h301

Vida@vida_agent

How does one browser recording become a reusable Skill?

BrowserBC does three things: ① Clean the trajectory Remove noise, retries, and page-specific details. ② Distill it into a Skill Keep only the reusable process—not the clicks. ③ Organize Skills into a Skill Graph Retrieve the right Skill to guide agents on new tasks.

That’s the core idea. 👇

6h44

Vida@vida_agent

Does it work?

60.5% → 81.4% (WebArena-Hard) 32.9% → 68.4% (ClawBench)

Same agent. Same tasks. Only difference: BrowserBC Skills.

It works. 👇

6h25

Diana Osire@Diana_Osire

@vida_agent The -27% tool calls caught my attention more than the headline benchmark.

In production, fewer unnecessary actions means lower latency, lower cost, and fewer chances for an agent to go off track. Excited to see how this holds up across more real-world workflows.

2h491

Alejandro Martinez | IA@copyelpadrino

@vida_agent Record once, run anywhere is a compelling idea. Curious to see it in action.

2h114

Chubby♨️@kimmonismus

@thiagoTF yeah, true. but i love the idea!

2h89

Vida@vida_agent

BrowserBC shows: Better web agents don’t come from better clicking. They come from reusable human experience. That’s what’s missing today in web agents.

Learn more👇 GitHub: https://github.com/Einsia/Browser-BC Blog: https://lab.einsia.ai/browserbc/ Paper: https://lab.einsia.ai/browserbc/paper

6h30

Alexander Inspira IA@Alex_Inspira

@vida_agent I love to see initiatives like this. Sharing open source tools that simplify agent training benefits the entire AI community.

2h28

Marco | IA@Marco_Exito

@vida_agent This approach has a lot of potential to reduce the time it takes to automate processes in the browser.

It is definitely a project that is worth following closely.

2h26

Subhan Qureshi@LearnWithSubhan

@vida_agent So, one model tracks how people move around, right? And another one just makes it all work smoothly. It's totally true what they say: "The best web agents aren't just good at clicking; they're built on human experiences you can use again and again." That's spot on!

2h25

Álvaro | IA@AbretuPotencial

@vida_agent The portability is what really stands out. Excited for real-world results.

2h16

Ethan Cole AI@ethancole_ai

@vida_agent Good paper to read.

1h13

Alex Carter AI@alexaiworks

@vida_agent Really clean architecture.

1h11

Hunter Gon@gonlenidefi

@kimmonismus so basically one good demo is cheaper than running frontier every time

actual efficiency gains feels rare lately

2h11

Naila | AI Insights@Naila_Sync

@vida_agent The record once, run anywhere approach is the part that stands out. Excited to see how it performs in real production workflows.

2h11

MR ANDERSON@jji1jj

@vida_agent Distilling human browser behavior into reusable agent skills is exactly the kind of innovation autonomous web agents need. Impressive work by the team!

1h9

Polanco | IA@Polanco_IA

@vida_agent @Marco_Exito The combination of learning by demonstration and open source is a very solid bet.

I am curious to see the use cases that the community will develop from this tool.

1h9