Himanshu Sharma

Half my job is building AI agents. The other half is promising they won't take over the world.

[Himanshu Sharma @Himanshush__]

@KhurramUdd93975 Missed 2 AM leads is a clear pitch. For the Gulf market, how does the agent handle Arabic and English in the same chat, and when does it hand off to a human?

[Himanshu Sharma @Himanshush__]

@slymn_ozcan Reading agent plans is a real use case. Do the plans come out as plain markdown files, or does the viewer need to handle something more structured like checklists and diffs?

[Himanshu Sharma @Himanshush__]

@a_vancin @boardyai Flagging Costco's own sites as clones is the false-positive that kills trust in this kind of tool. How did you fix it - an allowlist of known brand domains, or a better similarity check?

[Himanshu Sharma @Himanshush__]

@caiman__ 320 rule tests on a grind-size app shows where the judgment went. Which decision did the AI get wrong most often that you had to catch?

[Himanshu Sharma @Himanshush__]

@KleiAliajj That pause is also where you learn if the draft was worth sending. Do you track how often people edit before approving? The edit rate shows where the agent is still guessing.

[Himanshu Sharma @Himanshush__]

@ArmenDavtyan_ Two rejected reviews before a pass is the useful part. What did the reviewer catch that your own checks missed?

[Himanshu Sharma @Himanshush__]

@KanishkS3621 @expo Which of the Windows terminal bugs cost you the most time, and was it the Expo side or the tooling around it?

[Himanshu Sharma @Himanshush__]

@thecyberseckyle Of the 7, which one has someone outside your head waiting on it - a user, a client, a deadline? That might pick the winner faster than ranking how close each one is.

[Himanshu Sharma @Himanshush__]

@ArmenDavtyan_ Good that it cites sources. Did you check that each source actually says what the recommendation claims? Citations that don't support the claim are an easy miss on early runs.

[Himanshu Sharma @Himanshush__]

@MayankG22315198 Guarding against bad tool inputs is underrated. Are you doing schema validation, or checking whether the call makes sense in context?

[Himanshu Sharma @Himanshush__]

@AnonyfoxCom @X Three days learning what your own business does is the real cost of a usage limit. Was the gap missing logs, or no record of what the agent had decided?

[Himanshu Sharma @Himanshush__]

@lukaswarce The part that stays is whatever records state and gates risky actions. Curious which you'd bet on lasting longer: the state store or the approval step?

[Himanshu Sharma @Himanshush__]

@PaulBuilds00 Did this one come from a rejection, or did you catch it before submit? Wondering if review actually flags price prominence or if it only bites later.

[Himanshu Sharma @Himanshush__]

@siimh The nine months before the curve moves is the part nobody posts about. What moved first for you - impressions, clicks, or actual signups?

[Himanshu Sharma @Himanshush__]

Factory's Droid 1.0 lets the agent write a small program to coordinate tool calls instead of making them one at a time.

Factory says this cuts round trips and helps avoid skipped steps. I'm more interested in the second part. Fast is nice. Finishing the task is nicer.

[Himanshu Sharma @Himanshush__]

Building AI agents taught me a boring truth: getting one to do the thing is the easy part. Getting it to stop confidently lying to you is the job.

[Himanshu Sharma @Himanshush__]

@charles_maddock How is it different from the other agentic browsers?

[Himanshu Sharma @Himanshush__]

@achxvi But after how many tries?

[Himanshu Sharma @Himanshush__]

@yaosamo Haha this is so cool!

[Himanshu Sharma @Himanshush__]

@_annakulina That’s a good news! Glad it got found. Also started wondering how this could be translated to LinkedIn language. 3 lessons I learned when I lost my laptop 😂