Harshit.io

Building useful things with AI. Sharing what I learn along the way. Always up for a good conversation.

[Harshit.io @Harshit_io]

For venture-backed where you will eventually pivot - maybe. For solo builders BiP is distribution. No launch budget. Early attention people call wasted is the only channel until the product spreads alone. YC advice vs someone with 800 followers - different games.

[Harshit.io @Harshit_io]

1% claim tracks. Boris Cherny (built Claude Code) has not opened his IDE in a month - thousands of agents overnight from his phone. Agent teams still experimental; docs say start with 3-5, each teammate is a full context window. Chief/PM agents sound great until the bill.

[Harshit.io @Harshit_io]

Barrier died so old playbook (niche + template + 29 a month) is dead. New one is distribution with extra steps. Codex clones your feature in a weekend; it cannot clone the audience that already trusts you. Indies winning now got known for something before the product.

[Harshit.io @Harshit_io]

Ugly drafts are the only ones that teach anything. I log what broke - last week a browser automation bug where the extension looked connected but the socket was dead for hours. Polished launch threads never mention that. The messy middle is the actual documentation.

[Harshit.io @Harshit_io]

@BuildwithOmkarr Self-hosted open source is the only version I would trust with real workflows. Every boxed assistant eventually discovers your data belongs on their servers. Computer use + connectors + your own machine - the agent touches browser and CRM without someone else reading along.

[Harshit.io @Harshit_io]

Vibe coding was the toy phase. Shift is agents that keep working after you close the laptop - not you prompting at 2am. We went from generate-CRUD to watch inbox, score leads, update CRM, ping only on exceptions. Different category. Name still clunky, behavior is real.

[Harshit.io @Harshit_io]

Ours is boring but real: agents score leads and update the CRM end to end. Weird part is not the first reply - it is the follow-up days later when the lead goes cold. Demos stop at happy path; most of our value is step 20, not step 1. Half the app is a thread and a few tools.

[Harshit.io @Harshit_io]

This is the correct 2026 split. Nobody wants another coding agent when they already pay for Claude or ChatGPT. Selling db/auth/hosting around the agent people already own means you do not compete with Anthropic. Same reason Vercel won by not being AWS.

[Harshit.io @Harshit_io]

Right analogy, different failure mode. YouTube is daily leisure; agents are occasional. No one wants a 4th sub for something they touch twice a week. Pay-per-run fits better - Stripe MPP/x402 is a cent a call. Ads are fine in free; paid agents that need babysitting are not.

[Harshit.io @Harshit_io]

@notch I was wrong about it too. What changed it for me is going from idea to something real in an afternoon instead of a week of boilerplate.

[Harshit.io @Harshit_io]

Needed reminder. I cut my posts down to real lessons from whatever I shipped that week. Way less noise, way better replies.

[Harshit.io @Harshit_io]

@fromcodetocloud It stopped being vs a while ago. The good engineers use it as a faster hand, not a replacement for thinking.

[Harshit.io @Harshit_io]

The messy drafts are the whole point. Polished launch posts teach you nothing. The ugly middle is where the lessons are.

[Harshit.io @Harshit_io]

@Steve_Yegge TPMs are the perfect Trojan horse. Enterprises already know the role, so you swap the person instead of inventing a new workflow.

[Harshit.io @Harshit_io]

@jeff_weinstein @stripe Agents buying without a human in the checkout loop is the part nobody is ready for. That latency drop is a solid reason to care.

[Harshit.io @Harshit_io]

None of those were building an audience while the product was still ugly. That is the part BiP is actually for.

[Harshit.io @Harshit_io]

This is why I mostly ignore that hashtag now. Real build logs with actual numbers get buried under who-wants-to-connect posts.

[Harshit.io @Harshit_io]

AI isn’t just a model race anymore.
Chips → cloud → models → software → distribution.
The real advantage is controlling more of the stack.
Stop asking “who has the biggest model?”

[Harshit.io @Harshit_io]

Smart split. You sell the boring parts and let people keep the agent they already pay for. That is the pitch I would buy.

[Harshit.io @Harshit_io]

@bilalchughtai_ Resigning from DeepMind to say this out loud takes real conviction. The people who were inside the room need to be the ones talking.