Tobias Lutke posted on May 11 about Shopify's internal coding agent River. The detail that stands out is not what River can do - most agents can do similar things. It's what River refuses to do: direct messages. In 2026, an AI agent that won't slide into your DMs is apparently a design philosophy.
River works exclusively in public Slack channels. Every session is searchable. Any of the 100+ employees who watch River's work can join a session in real time - not as observers blocked behind a dashboard, but as active participants in an open channel. Lutke describes the effect as "Lehrwerkstatt" - the German term for an apprenticeship workshop, where learning happens through proximity to skilled work rather than formal instruction.
The restriction comes from organizational design, not from a technical limit.
Why this matters as a design choice
The default deployment model for enterprise AI coding agents is the opposite: private, audited, access-controlled. You get a chat window the width of your ambitions and the depth of your legal team's paranoia. The agent's work is visible to you and maybe to an IT security team reviewing logs after the fact - not to understand what the agent did, but to confirm nothing ends up in a discovery request. The rest of the organization sees nothing, learns nothing, and that is exactly how the vendor's enterprise tier is priced.
Shopify built River to invert that assumption. When River works on a task, it works where everyone can watch. Engineers who need to understand how the agent approaches a problem, what tools it uses, where it gets stuck, can learn that by observing - without anyone scheduling a knowledge transfer session.
The Lehrwerkstatt framing is the right analogy - assuming the theory survives contact with actual engineering culture. Traditional apprenticeship works because the junior engineer is physically present while the senior does real work, not because someone scheduled a knowledge transfer session and shared a Confluence page no one will read. Osmosis is a byproduct of shared space, not of formal training. River extends that to an entire engineering organization simultaneously. Whether 100+ engineers actually watch, or whether they mute the channel after day two, is a different question that Lutke's post does not answer.
The open question: sensitive data
The obvious objection is what happens when River handles information that shouldn't be in a public Slack channel - API keys, production credentials, internal architecture details that would make your CTO's blood pressure interesting. Lutke's post does not address this directly. No Shopify engineering post detailing those guardrails appeared in public records as of this writing. The announcement is considerably cleaner without those details.
Shopify has presumably solved this problem, either by restricting what River can access in the first place, by classifying which Slack channels are "public" within organizational norms (not externally public), or by training River to avoid exposing sensitive artifacts in thread outputs. Any large engineering org rolling out a similar model needs to work through the same question explicitly.
The broader organizational bet
What Shopify is betting on with River is that the productivity cost of full transparency is lower than the learning benefit of collective observability. That bet requires engineering culture that is comfortable with visible work - not just for the AI, but for the humans who interact with it. If your engineers are reluctant to work out loud, a DM-refusing agent creates friction rather than flow.
The organizations that benefit most from River's model are probably the ones that have already built strong working-in-the-open practices and just need an AI that respects those norms rather than undermining them. I have seen the opposite pattern play out on enterprise projects - agents tucked behind private integrations, generating knowledge at impressive speed that evaporates the moment the sprint ends and the contractor invoice is filed. If your company's default answer is "who needs to know about this?", River will feel like a punishment, not a feature.
Source: Tobias Lutke, Twitter/X, May 11, 2026, via Simon Willison's feed.