Thread by @ScottWu46

I'm often asked what programming will look like in 10 years or if it's even worth studying CS anymore. The future of software engineering is the single biggest question we think about in building Devin: 🧵

At its essence, programming is about telling computers what to do. Someday these computers may do everything themselves, but until then instructing them properly will be more important than ever.

The practical question today is less about replacement and more about how every engineer can be maximally effective.

AI has already made coding significantly more efficient synchronously - adoption among top engineers is close to 100% - but synchronous improvement is inherently capped.

Token generation is fast but not instant, and moreover a lot of software engineering is dependent on real-time events like running CI, spinning up the local dev server, or processing logs.

This means the next big unlock will come from building the interface and capability to delegate to multiple agents in parallel.

But delegation is not a single-shot process, and you often don’t know every implementation detail of what you plan to build when you first sit down. The multi-agent flow needs to support the ability to hand off and back on easily - describe the change, have the agent build the first pass, and then have it run the product locally so you can easily look and decide what’s next.

We’ve been building for exactly this experience in Devin 2.0. It’s a flow that takes time to adapt to, but we’ve fallen in love with it internally.

https://x.com/cognition_labs/status/1907836719061451067

It’s still raw, and we expect to make it much better and more intuitive over the next month, but we wanted to share it as soon as possible to get your feedback.

Try it out at and let us know what you think!app.devin.ai