Put AI in the group chat

The prevailing AI-centric working model can be pretty isolating. You bounce from agent to agent, making sure everything is on track. When you have a question, you ask an agent instead of bothering another human. You brainstorm with agents, get feedback from agents, and generally spend the majority of your day interacting with agents instead of humans.

It doesn't have to be like this. In fact, it really shouldn't be like this. Studies have shown that teams working with AI outperform individuals working with AI in creating exceptional outcomes. Unfortunately the most popular tooling today (commercial coding harnesses and desktop apps like Codex and Claude) is geared towards individual chat experiences (Claude does offer artifact sharing).

We think the benefits of this "multiplayer AI" are so indisputable that, as with cloud agents, it's a matter of when, not if. We've highlighted before that we're hearing of more and more teams building harnesses with collaboration and openness at the center: shareable sessions, Slack integration, open by default. But building a custom harness isn't for everyone. The good news is you don't need to.

For the baseline "multiplayer" experience, you need two things: an agent that can run in the cloud (not just on people's laptops) and integration with Slack or Teams. Slack is, in our opinion, where the magic happens. It's where much of the discussion around work is already happening (so there's built-in context) and by operating out in the open for everyone to see, a few people can teach your whole team how to work with AI by example. And to work effectively from Slack, an agent needs to be able to operate in the cloud.

You can easily buy this off the shelf: Devin, Cursor, Claude (Claude Tag), and Codex, as well as many others, have support for both of these things but vary widely in cost and ease of setup. Keep in mind that any cloud-based agent will only be as useful as the data and systems it has access to. Coding agents tend to be the easiest, because access, verification steps, and necessary guardrails are usually understood and already in place. For other data and systems, things may not be so easy. So be prepared to invest accordingly if you plan to go down this path.

"Managing a team of agents" is probably not where most people saw their careers going a year or two ago. Multiplayer AI can help infuse collaboration and humanity back into the work, and is a great way to supercharge AI adoption and leverage in your company.

Here's what we're reading this week.

The Multiplayer AI Manifesto

Multiplayer AI means people on the same team working with the same AI agents, in shared sessions and with shared learnings, instead of each person alone in a private chat.

Sergey Karayev from Superconductor lays out the case for multiplayer AI and how to do it effectively. It's a well-done website and an easy read (there's also a video if that's more your speed). Well worth your time if you're curious about multiplayer, considering buying a solution, or even considering building one.

Building a cloud prototyping environment

Jerry Di, Warp.dev intern, details why and how he built a cloud-based AI prototyping tool. There's so much to love about this, do yourself a favor and give it a read.

I wanted to build a system for designing that was persistent, reusable, and collaborative.

There's that word again: collaborative. Multiplayer. If you're building internal AI tools, are you building for single-player or multiplayer, and are you sure you picked the right one?

Towards self-driving codebases

Wilson Lin from Cursor with an update on their efforts to crack the code on long-running agent swarms for complex projects. The fact that the Cursor team, with nearly unlimited compute, has been working on this for 8 months hopefully provides some counter-balance to the litany of hyperbolic claims from social media. And better yet, we all get to learn from those months of experimentation.

One thing in particular that stood out to us: after experimenting with various agent roles, they landed on a simplified system with two basic roles (planner and worker) as the most effective. Matt Shumer independently landed on a similar system that he shared on X late last week, with "manager" and "implementer" roles.

There's lots of other interesting learnings as well, including explicitly loosening from 100% commit correctness and some generally useful tips on how to think about and interact with modern models and harnesses.

Treat the model like a brilliant new hire who knows engineering but not your specific codebase and processes.

Native is now the future of mobile at Shopify

Shopify announced they're moving away from React Native to Swift and Kotlin for their mobile apps. We love this because it's a concrete example of how AI is changing the tradeoffs in software development. Many of the constraints or challenges that shaped decisions five (or even two) years ago have shifted, and it may be worth revisiting those decisions in 2026.

What changed is that agents can now do enough of the implementation, translation, testing, and review work that it’s no longer the deciding factor it was in 2020.

The End of Code Review? Or an Opportunity to Rethink it?

Christian Kästner breaks down the cost of code review into a literal math formula, and then walks through how LLMs have changed each variable. It's a level-headed exploration that touches on the history of code review, its intrinsic and extrinsic value, and lands where we would land: change is coming, so let's figure out how to prepare and make the most of it.

So my main point is to stop asking whether and how code review in its current form can be saved, but to have an open conversation about what we are trying to achieve here and the costs and benefits of alternative models.

One thing to try this week

Find one skill to delete. It's not that skills are inherently bad. Far from it. It's just that the models keep getting better (and so do the harnesses) and some of the skills we created earlier this year aren't necessary anymore (or can be dramatically simplified). Take Wilson's advice and "treat the model like a brilliant new hire"; you may be surprised by the outcomes.

Thanks for reading!

— Claire + Zach