Harsh Kumar

Building @hanomiAI. CAD. LLM. Reasoning. Alignment
llm-pilled. code is artifact.
Maths major @iitroorkee.

[Sarvam @SarvamAI]

We are excited to announce that Sarvam is partnering with @PixxelSpace to power the AI backbone of India's first orbital data centre satellite.

This is a first for the country, with India-built AI models running on an India-built satellite and both training and inference happening directly in orbit, without any dependence on foreign cloud or ground infrastructure.

[Harsh Kumar @kumarsh0]

A useful pattern
- use several agent sessions [say A1...AN] to research/implement/debug a feature(s)
- get a showboat (agent's task specific proof of work) per session
- ask another agent to integrate all this to the main pipeline/workflow
- work with A1...AN to make sure their work is properly integrated
- this gives them more "final use case" context. Also gives them context on how their feature integrates with other incoming features

I unlocked insights that I couldn't see when working with each individual agent.

Showboats are a game-chaning concept from @simonw
https://t.co/qYCbYRJAW4

[Harsh Kumar @kumarsh0]

Yo blr people. Me and my friends are looking for a 3BHK apartment near Indiranagar, Domlur, HAL Road area

Ideally some place
- in a society with gym and pool (imagine that)
- not in a noisy environment
- cat friendly

If you are moving out of a nice place or know someone who is, do tell

Pls help @fmrbangalore @GruhamBot @BangaloreRoomi

[Ryo Lu @ryolu_]

overcooking

you've seen this: someone ships a dashboard that shows every number with a sparkline, every action has a confirmation modal, every empty state has an animated illustration and a tagline. individually each decision made sense to someone. together it feels like chaos. nothing is in focus.

that's overcooking. not one bad decision in isolation, but the accumulation of reasonable ones that no one said no to.

AI makes this worse as the cost of adding dropped to near zero. it can build a feature, even a whole new concept in minutes. so people do. and then they do it again. the thing that started with a clear purpose slowly becomes a collection of additions that are each justifiable but collectively incoherent.

the root problem is that most "new ideas" aren't new. they're repackaging of something that already exists at a more fundamental level. a new sticker on an old concept. it feels like progress because something changed, with a new word and skin – but the thinking didn't go deeper, it just duplicated itself into confusion.

the whole has a core. you feel it once you understand the whole system. everything in it are related and balanced. when you overload it, that gravity weakens. not because any one thing is wrong – but because attention is finite and you force it everywhere.

what we need aren't more tools that make more slop. it's seeing through the chaos, and returning to what the thing actually is, and cutting everything that doesn't serve that. that's harder now, not easier. because there's always something else you could add with one more prompt.

[Harsh Kumar @kumarsh0]

@CherryYadvendu @BangaloreRoomi @fmrbangalore yo ! this looks super cool. DM'ed you for more details

[Harsh Kumar @kumarsh0]

Looked into it. The approach is all reactive (hooks throughout the lifecycle, terminal bells). But :

- have to write different hooks per harness
- crucially, the agent never "knows" it's notifying you.
- lots of "false positives" as you are just guessing when the agent might need you

what if the agent can speak() ?

[Harsh Kumar @kumarsh0]

I work with a lot of agent harnesses simultaneously. And naturally go deep in one of the sessions (or drift away from all of them)

they finish tasks or wait for my input or hit a snag. I reply to them a lot later than I "should" have.

This is like you working with your most capable engineers in the same room and they just never tap on your shoulder. they just wait.

[Haoli Yin @HaoliYin]

https://t.co/N4W7BviNi0

[Harsh Kumar @kumarsh0]

Humans invented git for human collaboration reasons. It’s increasingly clear that Git-based workflows may not be suitable once we remove the human bottleneck from the flow of code.

[Harsh Kumar @kumarsh0]

We have good version control for the "code artifact" and almost nothing for the reasoning that got us there. Feels backwards

[Harsh Kumar @kumarsh0]

I work with a lot of agent harnesses simultaneously. And naturally go deep in one of the sessions (or drift away from all of them)

they finish tasks or wait for my input or hit a snag. I reply to them a lot later than I "should" have.

This is like you working with your most capable engineers in the same room and they just never tap on your shoulder. they just wait.

[Harsh Kumar @kumarsh0]

the hardest thing in my day to day workflow is getting everyone on the same page.

On both these fronts
- getting agents and me on the same page and
- extend to getting me, agents and other humans on the same page.

[Harsh Kumar @kumarsh0]

@drummatick cli ftw. Any good MCP can be made a better cli tool

[Harsh Kumar @kumarsh0]

@badlogicgames A simplification. Could be any feature that you (and pi) are not confident that they have the full picture. Take "creating animations like 3b1b" for example

[Harsh Kumar @kumarsh0]

Literally me. collaboration should no longer mean "you can take care of features X and Y, while I will get the infra running and you can benchmark and test while he completes the frontend integration"

This is all one person's job. If you understand it end to end, you implement it end to end. Period

Collaboration between humans makes sense when your own e2e understanding falls short or you wanna enrich it by arguing with other teammates

[Harsh Kumar @kumarsh0]

Let's say I feel the need of a github extension for pi. Me and pi decide that we wanna make it. We are aligned on goals. Now we want to know
- what all should a good and functional github extension must have ? (objectifying/planning)
- how are other similar extensions made these days (implementation)

How do you get these without web search ?

For context: I find claude's "Let me search how the community solves this problem" approach and follow up really helpful for these situations. Automatically enriches it's understanding

[Mario Zechner @badlogicgames]

i really want to people to understand that developer life is not a pokemon game. you do, in fact, not have to catch them all.

[Riley Brown @rileybrown]

Personally I don’t see the point of Cowork. I don’t even see the point of chatgpt or Claude anymore.

I either want to talk to a claw (agent with all my skills running on a persistent computer with NO guardrails in iMessage or telegram) or i want to use Claude Code or Codex app for coding.

[Harsh Kumar @kumarsh0]

We have good version control for the "code artifact" and almost nothing for the reasoning that got us there. Feels backwards

[Harsh Kumar @kumarsh0]

Something that bothers me is that when I think "how did I decide to route complex parts to a separate queue" the answer that I want isn't in the got log. The commit says something like "add tiered routing via rabbitmq" but that's not reflective of what all I thought

I am looking for the session where I spent 20 mins with claude to decide what to do and what not to. Why the current ways fits our stack and code and all