Harsh Verma

Previously @trycua @trynarrative

[Harsh Verma @kindaharsh]

what if your AI agent could call a human as a tool?

wired browser-handoff into a @browser_use agent doing a shopping checkout. when it hits the login wall or card form, it raises its hand. discord ping, I take over, agent resumes.

[James Murdza @jamesmurdza]

Human-in-the-loop undetectable browser-use!

[Harsh Verma @kindaharsh]

what if your AI agent could call a human as a tool?

wired browser-handoff into a @browser_use agent doing a shopping checkout. when it hits the login wall or card form, it raises its hand. discord ping, I take over, agent resumes.

[Burhan @BurhannKhatri]

No barriers with the Agent handoff by @kindaharsh, amazing work

[Harsh Verma @kindaharsh]

open source, drops into any browser-use agent as a custom action.

https://t.co/rzSVxPFPYO

full run (1 min): https://t.co/HhZ6t95dHO

[Harsh Verma @kindaharsh]

what if your AI agent could call a human as a tool?

wired browser-handoff into a @browser_use agent doing a shopping checkout. when it hits the login wall or card form, it raises its hand. discord ping, I take over, agent resumes.

[Harsh Verma @kindaharsh]

the agent decides when to ask — not the library. it has a `request_human_help(reason, done_when)` tool. when it hits something it shouldn't do (credentials, card), it raises its hand and waits.

I typed nothing outside the login + card form. agent did the rest.

[Neal Chopra @nealchopra]

A lot of people have been asking about our harness / approach - some thoughts:

1/ it’s fully open source on github!

2/ it is quite simple - and we think this is where harness engineering is heading. you no longer need elaborate scaffolding to force the model to reason in a prescribed way

3/ we initially included a verifier to check the executor’s work. it ended up being *more* accurate than the benchmark’s grader, but omitted it (you can't score above the ceiling set by the grader). we have a lot more to say on this.

4/ we were most excited by the performance uplift in sonnet (lighter model). it reflects a shift toward picking the model at the intelligence/cost pareto max for a task, not just the largest one. sonnet achieved near parity with opus in performance, while costing less than half.

[Harsh Verma @kindaharsh]

@kylejeong @jamesmurdza @daytonaio @browserbase instead of a manual hand-back button, it auto-detects completion via URL, element, content, or even LLM-based conditions - the agent resumes itself 👀

[Harsh Verma @kindaharsh]

I made browser-handoff, a tool that lets humans temporarily take over AI browser agents when human input is needed.

Built it while helping @jamesmurdza ship https://t.co/IFndgSwctr: sandboxed agents needed a way to log into Claude.

Demo runs in @daytonaio 👇

[Harsh Verma @kindaharsh]

@kylejeong @jamesmurdza @daytonaio @browserbase tried it - too laggy to complete a login, inputs barely worked 😅. Director is a product built around browserbase’s stream, browser-handoff is a library you attach to any playwright session

[Harsh Verma @kindaharsh]

@kylejeong @jamesmurdza @daytonaio @browserbase Cool! Though from what I can tell, Live View is the stream URL — browser-handoff handles the layer above: trigger detection, pause, notify, wait for completion, then resume. Couldn’t find that @browserbase does that natively — happy to be corrected if I’m missing something 👀

[Harsh Verma @kindaharsh]

Repo: https://t.co/rzSVxPFPYO

[James Murdza @jamesmurdza]

People keep telling me "Git was not designed for agents".

Git is SO good for agents: Commit, rebase, merge, repeat.

[Harsh Verma @kindaharsh]

@vague_anshul

[Harsh Verma @kindaharsh]

@vague_anshul hiiiii

[Brooklyn @vague_anshul]

X is cool.

but it’s 100x better when your timeline is full with people who code and build things.

if you’re into tech, AI, startups, product, design, development or programming.

say hi 👋

[James Murdza @jamesmurdza]

Claude Code is free right now!

[James Murdza @jamesmurdza]

Thanks to:

@pluuto19
@kindaharsh
@burhannkhatri

for their great open source contributions!

[Harsh Verma @kindaharsh]

@jamesmurdza

[James Murdza @jamesmurdza]

I made an open source version of Claude Code for web. And it’s way better:

Six different agents
Upload and view files
Fully featured terminals
Nested subagents
Easily rebase and merge
It's open source

Every agent and branch runs in a @Daytonaio sandbox.