I'm Aloke, first engineer and head of the coding team at Warp. AMA! 🔥

@thinkverse Great question. We're still in the nascent stages of scoping out LSP, my guess is we'll start by supporting LSP out-of-the box for the top ~10 languages and then follow up with support for custom LSPs for the language of your choice.

@gabe Thanks for the support Gabe!

The primary way I work now is to have the code review pane open alongside a session where Warp is writing code. I can see what the agent is doing and then review the code and have it address my feedback.

For simple changes, I just edit the code inline in the code review panel or I'll pop open the file in its own tab for slightly more complicated changes.

I find this to be the best way to write code with an agent. You can quickly start a task via a prompt, see the changes the agent has made in the code review panel, easily course-correct it by attaching the diffs as context, and directly edit files in Warp to make quick changes yourself.

@mountroot Great question! One of the biggest early challenges was using Rust to create our UI framework.

Performance has always been really important to us at Warp, and we knew we wanted to render on the GPU so actions like displaying tons of output from long-running commands wouldn’t become a bottleneck.

Back then (about 5 years ago) there wasn’t an off-the-shelf library that fit our needs, so we had to build our own. The tricky part is that the usual ways of modeling UI frameworks in object-oriented languages don’t map cleanly to Rust.

If you’re curious about the details, I wrote up the design decisions we made in this blog post: https://www.warp.dev/blog/why-is-building-a-ui-in-rust-so-hard

@mikekerzhner Great questions!
1) The lethal trifecta is serious and something no AI company has really solved yet IMO. The key piece is whether the agent has the ability for "external communication" (in reality access to the internet in some way). The way we plan to address this is to denylist commands that could use the internet (such as curl and dig) so users always have to confirm it's safe before it runs. For future products we build that may run on remote machines, we're going to introduce the concept of network isolation where you can give an agent full network access, access to just app.warp.dev, or a domain allowlist.

2) For now we think we can get the best quality from SOTA coding agents. We may train our own models if there are specific use-cases where we think we can outperform claude or GPT5, but for most coding use cases I think the SOTA models shine and it's a distraction for us to try to build our models.

@gsmbk Thank you so much for the kind words! We don't have a plan to support a mobile app just yet, but hearing feedback about why you'd find it useful helps us a ton in determining whether to prioritize it at some point in the future.

@moelove we should def support Vim keybindings for the file tree. I’ll make sure that gets fixed

@abod_rehman Thank you so much! We invest a lot of time in the overall product experience of Warp to make sure it's an awesome experience. The combo of great UX + SOTA agents is what makes Warp so good in my opinion.

@beginners_blogFirst, it's essential to establish a clear definition of 'full autonomy.'

I am developing a service, "Vibe Coding," that aims to achieve this.
Our approach involves an AI that, based on user specifications, autonomously selects from four backend architectures and configures everything down to the communication protocols for each function.

While this achieves autonomy in performing its function, I also think it might be difficult to classify it as truly autonomous. True autonomy should have no limitations and must account for factors like optimization and code elegance. In reality, limitations exist, and optimization is clearly still a necessary factor.

Ultimately, it seems a developer will always need to be intimately involved, stepping in at the stages where the AI agent falls short. That said, the core challenge for developers will be to determine just how far an AI agent can go in delivering results that satisfy users without needing human intervention. And aren't many teams in this field striving to constantly expand that range of satisfaction? 

@cameronraysmith No plans on making session sharing fully functional on mobile (yet)

But we are actively working on adding our AI features in session sharing. We also are building an SDK that can run on remote machines where you can use session sharing to quickly jump in and see what the agent on the remote machine has done.