Cline: the open source coding agent that doesn't cut costs

**Alessio** (0:05)
Hey everyone, welcome to the Latent Space Podcast. This is Alessio, partner and CTO Decibel, and I'm joined by my co-host, Wix, founder of Small AI.

**Wix** (0:12)
Welcome, welcome. And today in the studio with a nice two guests from Cline, Pash and Saoud.

**Saoud Rizwan** (0:19)
That's right. Yes.

**Alessio** (0:21)
You nailed it. Let's go.

**Wix** (0:23)
I think that Cline has a decent fan base, but not everyone has heard of it. Maybe we should just get it like an upfront, like what is Cline maybe from you, and then like you can modify that as well. Yeah.

**Saoud Rizwan** (0:35)
Cline is an open source coding agent. It's a VS Code extension right now, but it's coming to JetBrains and NeoVim and CLI. You give Cline a task and he just goes off and does it. He can take over your terminal, your editor, your browser, connect to all sorts of MCP services, and essentially take over your entire developer workflow. It becomes this point of contact for you to get your entire job done essentially.

**Wix** (1:02)
Pash, what would you modify? Or what's another way to look at Cline that you think is also valuable?

**Pash** (1:08)
Yeah, I think Cline is the infrastructure layer for agents, for all open source agents, people building on top of this agentic infrastructure. Cline is a fully modular system. That's the way we envision it and we're trying to make it more modularized so that you can build any agents on top of it. So with the CLI and with the SDK that we're rolling out, you're going to be able to build fully agentic systems for anything, not just coding.

**Wix** (1:36)
Oh, okay. That is a different perspective on Cline that I had. So, okay, let's talk about coding first and then we'll talk about the broader stuff. You also are similar to Adr. I don't know who comes first in that you use the plan and act paradigm quite a bit. I'm not sure how well known this is. Like to me, I'm relatively up to speed on it.
But again, like maybe you guys want to explain like why different models for different things.

**Saoud Rizwan** (2:03)
Yeah, I'm going to take the cred for coming up with plan act first. Cline was the first to sort of come up with this concept of having two modes for the developer to engage with. So just in like talking to our users and seeing how they use Cline where it was really only an input field, we found a lot of them starting off working with the agent, coming up with a marked on file where they asked the agent to put together some kind of architecture plan for the work that they want the agent to go on to do. And so we would find that people just came up with this workflow for themselves just organically. And so we thought about how we might translate that into the product. So it's a little bit more intuitive for new users who don't have to kind of pick up that pattern for themselves and can kind of direct and put in guardrails for the agent to adhere to these different modes whenever the user switches between them. So for example, in plan mode, the agent is directed to be more exploratory, read more files, get sort of understanding and fill up its context with any sort of relevant information to come up with a plan of attack for whatever the task is the user wants to accomplish. And then when they switch to act mode, that's when the agent gets this directive to look at the plan and start executing on it, running commands, editing files. And it just makes working with agents a little bit easier, especially with something like Cline, where a lot of the times people's engagement with it is mostly in the plan mode, where there's a lot of back and forth, there's a lot of extracting context from the developer, asking questions, what do you want the theme to look like, what pages do you want on the website, just trying to extract any sort of information that the user might not have put into their initial prompt. Once the user feels like, okay, I'm ready to let the agent go off and work on this, they switch to act mode, check auto approve, and just kick their feet up and get coffee or whatever and let the agent get the job done. So yeah, most of the engagement happens in the plan mode and then act mode, they kind of just have a peripheral vision into what's going on, mostly to course correct whenever it goes in the wrong direction. But for the most part, they can just rely on the model to get it done.