Why Autonomous AI Agents Are Essential for Vibe Coding Websites (James Dooley Interviews Kasra Dash)

James Dooley (0:00)

Agentic orchestration for vibe coding new websites. There's many people now that's scaling out new websites, whether they're exact match domains, partial match domains or new branded websites, and the vibe coding these using AI orchestrators. And I think the most important part of this is using the AI harnesses and AI orchestrators to do it for the quality control, as opposed to just asking Claude code directly, can you build me X, Y and Z?

Today, I'm joined with Kasra Dash, who we're going to talk a little bit more about the agentic orchestration of vibe coding. So for anyone that doesn't know what vibe coding is, can you just briefly explain what vibe coding is for a new website?

Kasra Dash (0:47)

Vibe coding is basically when you get Claude or ChatGBT or any LLM really, to code you an actual website. I think it's become a lot more popular in the last maybe one or two years, because it's become so accessible for people that are non-techie, non-developers to just go in, give ChatGBT an idea or a spec sheet, and basically it will actually start vibe coding your idea. A lot of people have actually, what they've started to do is they might actually hire a developer on Fiverr or wherever at work, create a spec sheet, and then they will hand that spec sheet over to an actual vibe coding platform.

James Dooley (1:34)

So with regards to the agentic orchestration of it, obviously you're using Kodus at present as being the AI harness. And that third party kind of AI harness can then go and connect in to Codex.

So if you think that Astra is better than, let's say, Fable 5.1, you could then switch between the LLMs very easily. But can you explain a little bit of why the agentic orchestration, and having that orchestrator that's able to deploy different agents in parallel with each other, why that's important as opposed to just using, let's say, Claude, and only using Claude code?

Kasra Dash (2:13)

Sometimes just using Claude or Fable 5.1 for everything, it's almost like taking a tank to a race. You don't necessarily need a fighter tank or a fighter jet. Sometimes all you need is just a pair of running shoes to win the race.

I think that's where a lot of people end up burning a load of tokens. So when you have an orchestrator, it will read your prompt, so whatever you've given it, and it will break that prompt down into subsections. So it's like, right, okay, Kasra needs images for his website. Well, Claude can't create images, so I'm gonna give that task to Image 2.0.

Then it might read it further down. It might say, okay, Kasra also wants us to scrape 50 websites. Well, I'm definitely not gonna give that task to Fable 5.1 because scraping with Fable 5.1 is just way too expensive. I'm gonna give that to maybe a Chinese model like Kimi K. So it understands your prompt, and then it starts planning out which AI agent is actually the best for that given task.

James Dooley (3:23)

So if you're using that AI orchestrator and it's able then to kind of go, okay, I'm gonna use ChatGPT, I'm gonna use Claude, or I'm gonna use Quen or Kimi, and it distributes to them different LLMs. Are you also distributing different tasks to the different agents? So let's say they're gonna do the build out, you're gonna have a schema agent, you're gonna have a topical map agent, you're gonna have a content brief agent, a copywriting agent, an internal linking agent. Do they all run at the same time? Does your agentic orchestration, is that kind of setting all out right from the word go to deploy those parallel agents, survive code and get to the end result faster?

Kasra Dash (4:09)

Yeah, so I might have like 25 different tasks running at once, all trying to push different things. So you mentioned, for example, schema. I would never give Fable 5.1 or Astra 6 schema because schema's basically already been predefined. You already know what correct schema is, and you already know what incorrect schema is.

All you basically need to do is give that documentation to, let's say, for example, Kimmy, and it's more than capable to read the documentation and be like, right, okay, I get this. I'm going to start coding and understanding what kind of schema we actually need on this website. Whereas if I'm creating something from scratch that's never been done before and there is no documentation to say, that's when I would probably give it to potentially an Astra or a Fable 5.1, because it's something from scratch and they might need to do a little bit of research for themselves and a little bit of deep research.