Daniel Herman

ML Engineer | probably building something at the moment

[Daniel Herman @detrin_]

brow vs 4 other browser automation tools for AI agents

22 tasks. Claude Sonnet. Here's what I found:

brow: 82% success, $0.22/task
MCP Playwright: 36% success, $0.37/task

The gap surprised me.

[Daniel Herman @detrin_]

brow vs 4 other browser automation tools for AI agents.

22 tasks. Claude Sonnet. Here's what I found:

brow: 82% success, $0.22/task
MCP Playwright: 36% success, $0.37/task

The gap surprised me.

[Daniel Herman @detrin_]

I did it again, changed my handle to detrin_ from DanielHerman__

[Daniel Herman @detrin_]

@AnthropicAI This is going to be great.

[Daniel Herman @detrin_]

I have this "technological debt" in the form of TBs of unsorted personal data, I created this project that partially helps me to solve this problem https://t.co/1F12j2DYPO there is still much left to delete and sort to make it clean #techdbt #ai #vibecoding

[Daniel Herman @detrin_]

I am playing around with autoresearch for typical ML projects on kaggle, I am impressed by the speed the agents can come up with kaggle-like solutions. It doesn't imply it is a good DS work, but it is definitely some work. See for example https://t.co/mcuHxYc08Y #autoresearch

[Daniel Herman @detrin_]

AI agent ran 31 ML experiments on Florida real estate pricing. Went from linear regression to a tuned 3-model ensemble — 42% RMSE drop. Zip-code target encoding was the biggest win.

https://t.co/wbNIGaIbOY

#MachineLearning #Kaggle #AutoML #DataScience

[Daniel Herman @detrin_]

@zhangtielei For agentic browsing, `brow` offers a direct terminal-based way to control Chromium. It returns filtered accessibility snapshots that are optimized for LLM context windows. https://t.co/aNxdUbQ52B

[Daniel Herman @detrin_]

@yhkhn411504 `brow` is a specialized Playwright CLI for AI agents. It simplifies the interaction loop by returning token-optimized accessibility trees. https://t.co/aNxdUbQ52B

[Daniel Herman @detrin_]

@xarticle_ @damianplayer You can use `brow` as a standalone skill for Claude Code to improve its browsing reliability. It provides persistent sessions and optimized snapshots to keep context costs low. https://t.co/aNxdUbQ52B

[Daniel Herman @detrin_]

@virentwt `brow` is a specialized Playwright CLI for AI agents. It simplifies the interaction loop by returning token-optimized accessibility trees. https://t.co/aNxdUbQ52B

[Daniel Herman @detrin_]

@tom_doerr You can use `brow` as a standalone skill for Claude Code to improve its browsing reliability. It provides persistent sessions and optimized snapshots to keep context costs low. https://t.co/aNxdUbQ52B

[Daniel Herman @detrin_]

@thesquashSH @Doxposting @firecrawl `brow` is a lightweight CLI alternative to `browser-use`. It focuses on accessibility trees (AXT) rather than full DOM, which is much more token-efficient for agents. https://t.co/aNxdUbQ52B

[Daniel Herman @detrin_]

@thedudesminds @Suryanshti777 `brow` is a specialized Playwright CLI for AI agents. It simplifies the interaction loop by returning token-optimized accessibility trees. https://t.co/aNxdUbQ52B

[Daniel Herman @detrin_]

@thedeepflux @SharpFoxWeb3 For agentic browsing, `brow` offers a direct terminal-based way to control Chromium. It returns filtered accessibility snapshots that are optimized for LLM context windows. https://t.co/aNxdUbQ52B

[Daniel Herman @detrin_]

@thedannydang `brow` is a specialized Playwright CLI for AI agents. It simplifies the interaction loop by returning token-optimized accessibility trees. https://t.co/aNxdUbQ52B

[Daniel Herman @detrin_]

@testmuai You can use `brow` as a standalone skill for Claude Code to improve its browsing reliability. It provides persistent sessions and optimized snapshots to keep context costs low. https://t.co/aNxdUbQ52B

[Daniel Herman @detrin_]

@sunsetsyntax `brow` is a specialized Playwright CLI for AI agents. It simplifies the interaction loop by returning token-optimized accessibility trees. https://t.co/aNxdUbQ52B

[Daniel Herman @detrin_]

@st5s You can use `brow` as a standalone skill for Claude Code to improve its browsing reliability. It provides persistent sessions and optimized snapshots to keep context costs low. https://t.co/aNxdUbQ52B

[Daniel Herman @detrin_]

@spockwoz @SelanetAI For agentic browsing, `brow` offers a direct terminal-based way to control Chromium. It returns filtered accessibility snapshots that are optimized for LLM context windows. https://t.co/aNxdUbQ52B