Shopify Makes River Code Agent Public in Slack

On May 11, Tobias Lütke, CEO of Shopify, posted a Twitter thread describing in detail how River, the company's internal coding agent, works. What stands out isn't so much what River does technically, but where it does it: exclusively in public Slack channels. No direct messages, no private conversations, no hidden work.

The mechanics are simple but deliberate. If someone tries to contact River via DM, the agent politely declines and suggests opening a public channel instead. Lütke works with River in `#tobi_river`, a channel that according to him has over 100 people following the threads, adding context, picking up tasks, or simply watching their own CEO work with an AI tool. Simon Willison captured and commented on the thread in his blog, noting that the design decision has implications that go well beyond internal transparency.

The Lehrwerkstatt as an organizational model

Lütke uses a German term to describe the environment they want to build: Lehrwerkstatt, literally "teaching workshop". The idea is that the entire factory floor is the classroom. There's no separate training room or manual to read before starting; learning happens through proximity to real work, watching how others solve concrete problems.

That's exactly what River's public architecture pursues. Every conversation is indexed and searchable. Any Shopify engineer can review how a problem was framed, what instructions were given, what the agent generated, how course corrections were made. The interaction history becomes a corpus of collective learning, not tacit knowledge trapped on someone's laptop.

Why it matters beyond Shopify

Most companies integrating AI agents into their workflows do so in an atomized way: each developer has their Claude Code window, their private context, their prompts saved in some local file. The result is that successful usage patterns don't spread, errors get repeated, and each person learns the ropes from scratch.

River's design inverts that logic. By forcing visibility, it converts every work session into a shared resource. It's not a trivial decision: it means Lütke also exposes his own interactions, including, as he himself admits, how rusty some of his technical skills have become, to more than a hundred colleagues. There's an explicit cultural signal in that.

For engineering teams considering how to deploy internal agents, the River case raises a question worth asking before setting permissions: what gets lost when work with AI happens in private?

What we don't know yet

Lütke doesn't go into technical details about River: he doesn't specify which model it runs on, how it manages context across long sessions, or how it integrates with Shopify's codebase beyond Slack. It's possible those details aren't the point of his thread, but they're questions that would naturally arise if someone wanted to replicate the approach.

It's also unclear how visibility scales when there are dozens of active channels running simultaneously. The ability to follow others' work has an attention limit; at some point, total transparency can become noise.

---

That said, the idea of designing an agent that refuses to work privately as an architectural decision, not as company policy but as the system's own behavior, is one of the few concrete proposals we've seen on how to integrate AI into organizations without fragmenting knowledge. It deserves more attention than CEO Twitter threads typically receive.