Swapnil

Deep dives into AI tools, agents, and LLMs. Honest takes on what's signal vs. noise.

[Swapnil @t_swap]

If you’ve been doom-scrolling AI Twitter, you’ve definitely seen the hype around multi-agent swarms at some point. Someone claiming “I built a 12-agent swarm that codes, writes, and launches SaaS products while I sleep.” Meanwhile, in your actual workflow, you’re still copy-pasting between three different chats and praying nothing explodes.

Honestly, I started to get FOMO for not running enough parallel agents. So I set out to find: do people actually use multiple agents simultaneously in a meaningful way today? And more importantly: outside of coding and tech demos, is this “agent team” thing real, or just a fancy word for “I opened three tabs”? Not the influencer hype but the ground reality. So I spoke to real people, real developers and researched some multi-agent capabilities available in the popular coding tools today. If you've been curious about it as well, then this one is a gold mine for you - https://t.co/okuigUMe4n

[Swapnil @t_swap]

9/9 A much better thing to organise now (as was always) is a SELL-A-THON. Give people an already built product and ask them to sell it and earn their first dollar in 24 hours. I think that's a much more valuable skill today than continuing to build just another vibe-coded app.

[Swapnil @t_swap]

1/9 One of my friends recently asked me if I wanted to join him for one of the weekend AI Hackathons. It's been a while since I attended a hackathon.

[Swapnil @t_swap]

8/9 This is not a diss on hackathons. I've made many friends and even built my first-ever product in a hackathon. But I do think they need to be rethought to stay relevant today.

[Swapnil @t_swap]

@alliekmiller Curious - when you say you're not worried about token with this or Fable - why is that? Is it because the results are so good and they justify the cost? Or do you have a special discount?

[Swapnil @t_swap]

What makes AI agents fascinating and gives them the "wow" factor is also the same thing that LLMs inherently lack and makes them look dumb despite all their knowledge - MEMORY.

Memory is what makes AI agents look alive but memory also often gets confused with two things it isn't. It isn't the context window, which disappears the moment a session ends, and it isn't a skill, which stays static across runs. Memory is the part that changes every time, because it's the record of what actually happened: which fixes failed, which decisions got made and why.

This episode is the follow-up to my loop engineering series where I dive deeper into the fascinating concept of Agent memory. I go through three patterns for building this, starting with a plain Markdown file, through task trackers, up to structured stores for when a loop truly outgrows a flat file.

I also cover where memory breaks and what you should do about them. If you're building agent loops that are supposed to get smarter over time, you need to listen to this. - https://t.co/iaiCzORnlq

Fair warning - this post is only useful if you're either building your own agents or are interested in knowing the agent internals.

[Swapnil @t_swap]

@petergyang This can be vaalu different for different domains. Can you give some specific examples?

[Swapnil @t_swap]

And as was bound to happen, Reddit had a field day with this.

[Swapnil @t_swap]

I'm sure Beth Galetti had good intentions when she shared this post - https://t.co/PdVGFwpvb5, and she meant to encourage all the non-tech people who are intimidated by AI to take the leap and experiment more.

[Swapnil @t_swap]

There is undeniably a huge productivity unlock waiting for people who are willing to experiment and vibe with AI. But we need to be careful not to position vibe coding to be at the same level as coding. The lines of code should never have been mentioned because it's not about that. If she had simply stuck to talking about the huge efficiency gains she got from the little tools she was able to build on her own, I'm sure that'd have landed better.

[Swapnil @t_swap]

While every company is trying to become AI-native, few are getting it right. Companies can take a page from Shopify's playbook https://t.co/UF6PvCT2Om

[Swapnil @t_swap]

Why would you want to do that when you're already getting them cheap, subsidized by big frontier labs that have 100s of scientists and $$ building this. Doesn't it make more sense to build consumer apps using existing LLMs that solve real problems? Or as @ylecun said, clean his bedroom?

[Swapnil @t_swap]

@alliekmiller It's funny that most of the people in the thread think it has to do with cost. Limiting context window is a technical limitation and affects how accurately LLM will respond to your query.