shivangi gupta

[shivangi gupta @Shivi8813]

@Sushilk91 @neilshroff Not to boast but i love the shared notes part and the timeline feature where we jot down something special from day/week ; just images or memories! 🧿

[shivangi gupta @Shivi8813]

@shobhitic The work that they do(using their services proactively for ~yr now), seriously need better marketing team. These kind of events are totally unnecessary to grab people's attention.

[shivangi gupta @Shivi8813]

@shobhitic Super irritating FNP reminders are!

[shivangi gupta @Shivi8813]

@Sushilk91 Do you think harness could eventually be a product itself? Or remains as a commodity?

[shivangi gupta @Shivi8813]

@ISpiriTualist OMG! 🤩

[shivangi gupta @Shivi8813]

@Sushilk91 How do i access the tool?

I like the name Velora! 🙃

[shivangi gupta @Shivi8813]

Totally. I have been seeing this in day-2-day trivial tasks too!

Speed and the efficiency is directly proportional to how good and crisp context is given. 👩‍💻

[Sushil Kumar @Sushilk91]

Model choice matters here, but in my own runs, most of the savings came from something much less exciting: context management.

A long-running agent keeps pulling its entire history into every next step, even when most of it has stopped being useful. In one of my setups, around 70% of the spend was cache reads.

What worked much better for me was giving every worker a fresh context and passing only what it actually needed: the task, a small state file and the relevant repo context.

It cost less, but the workers also did better. They stayed focused, drifted less and didn’t get confused by decisions from hours ago.

I do this using yoyo. For bigger tasks, I either run fresh workers directly or ask the main agent to spin up its own subagents for focused pieces of work.

Those subagents don’t always need the smartest model. Once the task is clear, a smaller and cheaper model is often enough.

[Sushil Kumar @Sushilk91]

"stop prompting, start engineering loops" is the only AI coding advice I actually believe right now.

spent the weekend rewriting my agent harness. 40% reliable task completion → 90%+. same model. same tools. nothing changed except the loop around them.

what moved the needle:
- explicit exit conditions + a separate critic LLM grading the executor
- progress.md on disk instead of in the context window
- stagnation detection that kills infinite loops before they burn budget

[shivangi gupta @Shivi8813]

@shobhitic On mobile data, it works

[shivangi gupta @Shivi8813]

@shobhitic Negative civic sense! 🥶

[shivangi gupta @Shivi8813]

@shobhitic 🫣 nooooo

[shivangi gupta @Shivi8813]

@shobhitic The "i know founder" power!! 🫡

[Sushil Kumar @Sushilk91]

US just banned Claude Fable 5 and Mythos 5 from all foreign nationals. sovereign AI isn’t a buzzword anymore. whoever owns the models owns the leverage. the US knows this. China knows this. India is still forming committees.

our entire AI budget is ₹10,300 crore. we spent ~₹1 lakh crore on Ayodhya and Kashi. ₹1.35 lakh crore on the 2024 election. most expensive election in human history. money was never the problem.

R&D at 0.64% of GDP. for a decade. china is at 2.43%. US at 3.48%.
india’s big GPU program: 38,000 GPUs. china ordered 2 million+ H200 chips this year. 38,000 vs 2,000,000.

73% of our top AI researchers leave and don’t come back. why would they.

we threw money at marble and sandstone while the world was building compute and research labs. india won’t be remembered as a country that couldn’t do AI. we’ll be remembered as the one that just chose not to.

[shivangi gupta @Shivi8813]

@Sushilk91 Itna complex! L 😳

[shivangi gupta @Shivi8813]

@Sushilk91 Wow! That's a great news to empower the AI solutions even more.

[shivangi gupta @Shivi8813]

@Sushilk91 Whattttttt 🫩

[shivangi gupta @Shivi8813]

@Sushilk91 I have been using copilot with claude models and suddenly there is no more premium models on the list. i was wondering what changed. Thanks for sharing the post in details.

So afterall this circle, hiring people is cost effective than the models 😹

[Sushil Kumar @Sushilk91]

GitHub Copilot just switched to token-based pricing, and people are posting bills that have increased 10x overnight.

Everyone is blaming the pricing.

I run coding agents all day while building Farfield, and most of that bill is self-inflicted.

The model didn’t get more expensive. Your harness did.🧵