Shopify's River AI Agent Works Only in Public

Everyone covered the numbers. I want to talk about the rule.

When Toby Lütke posted about Shopify's internal AI coding agent, River, the stats were genuinely wild—5,938 employees using it across 4,400+ Slack channels in a single 30-day stretch this spring, 1,800 pull requests opened in one week, roughly one in every eight merged PRs now coming from River. (Note: these figures come from Lütke's post as cited by AI strategist Nate B. Jones; I haven't independently verified the precise timing or the post URL.) Those numbers got screenshot-and-retweeted into oblivion. Fine. Big numbers are fun.

But here's the part that actually got me: River cannot run in a DM. Technically cannot. If you try to talk to River in private, nothing happens. Every conversation an engineer has with this agent exists in a public Slack channel, scrollable by anyone on the team. That's not a privacy oversight or a Slack API limitation. That's a deliberate design constraint, and it's doing something way more interesting than shipping pull requests.

Jones, who covers AI strategy, calls what happens when you remove that constraint the "apprenticeship gap." I think he's named something real.

The Problem That's Not a Tooling Problem

Here's a thing that's probably true about your workplace right now: your colleagues are running AI workflows you've never seen. They've got prompts that reliably cut a two-hour task down to fifteen minutes. They figured out exactly how to load context into Claude so it doesn't hallucinate the edge cases. They built a little vibe-coded script that does something useful and they're the only one who knows it exists.

Jones describes talking to people inside Amazon who mentioned—anecdotally, so treat this as illustrative rather than audited—that there are reportedly six, eight, even ten different internally-built tools solving the same problem because each builder didn't know the others existed. Individual employees compounding. The organization standing still.

If you're early in your career right now, you already know exactly what I'm talking about from the other direction. You've got workflows your manager has never seen. Which, fine, those are yours. But you've also never seen theirs—and that's the half of the deal that actually costs you something. For most of human history, Jones argues, you learned skilled work by being physically near skilled people. You watched how a senior person framed the problem, not just solved it. You absorbed taste by proximity. That transfer is now getting routed through private chat windows where nobody can see it.

What Polanyi Has to Do With Your Slack

Jones reaches for a concept from philosophy of science called Polanyi's paradox—the idea that we know more than we can tell. The tacit stuff, the intuition, the judgment that experts can't fully articulate even when they try. He uses manufacturing as a case study: entire generations of machinists nearing retirement who carry knowledge literally in their fingertips that cannot be fully rendered into documentation or training data. Jones mentions illustrative examples—a craftsperson who paints racing stripes on Rolls-Royces, a machinist somewhere in the Pacific Northwest who knows how to quality-test a specific Boeing component—as the kind of singular expertise that defies capture. (These appear to be rhetorical illustrations in Jones's telling, not citations with traceable sources.) Product managers have tried to bottle this knowledge into ML systems before those workers retire, Jones says, and the consistent finding is: you can get close, you can approximate, but you can't get there.

Software has the same problem in a different costume. The senior engineer who knows when to reject an AI output—immediately, almost reflexively—is carrying tacit knowledge about quality that looks like magic to someone who's never watched them work. The correction that turns a plausible-but-wrong model response into something actually usable: that's the craft. And it's been disappearing into private browser tabs.

This is where Shopify's constraint becomes more than a Slack policy. River being public means that when a senior engineer is debugging a tricky problem with the agent, everyone on the team can scroll back later and see the whole thread. Not just the merged PR. The session. Where the engineer gave up on the first approach. What context they pasted in. What the model got wrong and how the engineer caught it. What got rejected and why.

That scroll-back is the apprenticeship. Anthropic's internal Claude deployment in Slack is trying to solve something adjacent—the gap between engineers who ship faster with AI and the rest of the team waiting for them—but Shopify's approach makes the reasoning visible, not just the output.

What "Public AI Work" Actually Looks Like

Jones breaks down four things worth making visible, and I want to translate this out of framework-slide language because it's actually practical: the task (what were you actually trying to do?), the context (what did you load in, and importantly, what did you leave out?), the interaction (how did the back-and-forth actually go—what did you push back on?), and the review (what did you accept, what did you manually rewrite, and why?).

Picture a junior engineer who just joined the team scrolling back through an older River thread on a Monday morning. They're not reading a prompt library. They're watching a senior engineer realize mid-thread that the model is solving the wrong problem, redirect it, and then—this is the part that doesn't exist anywhere else—reject the second answer too because it missed a constraint that matters for this particular customer. That moment of "no, still wrong, here's why" is the whole lesson. It's the difference between knowing that AI makes mistakes and knowing how a good operator catches them. Jones is explicit about this: "The most valuable part of AI work is rarely the prompt. It's the surrounding habit."

A prompt library can't give you that. A prompt library is a recipe without the cook's intuition about when to deviate from it.

Jones also notes—and this tracks with what I see when people watch experienced AI users work—that one of the most surprising things is how often and how fast good operators say no to the model. You build that reflex by watching it, not by being told it exists.

The Senior People Problem

Here's where Jones gets uncomfortable in an interesting way. Senior people have the most valuable AI judgment and the least visible process. They're pressure-testing strategy decks with agents, comparing scenarios, finding weak assumptions in roadmaps—and all of it is happening offstage, increasingly with AI doing heavy lifting that the rest of the org never sees.

According to Jones, Lütke has reportedly treated this as a leadership responsibility: putting his own River interactions in public channels, letting team members ask questions of his agent, opening up his process to scrutiny. Jones describes it as "a little bit chaotic" but deliberate—Lütke is using the visibility to socialize how he wants people to think about using AI, not just that they should use it. (This behavior is attributed to Lütke in Jones's video; I'm treating it as Jones's reported characterization, not independently verified.)

The Stripe model of agentic engineering—1,300 PRs weekly from AI agents—shows what raw throughput looks like when you optimize for output. Shopify's River story seems to be asking a different question: what does throughput look like when you optimize for institutional learning at the same time?

Does This Actually Scale?

Okay, here's where I'm going to be honest about my skepticism, because I think it matters. Jones's argument is compelling at the team level. One declared public channel, a pinned message at the top explaining the norms, senior people running real work where junior people can watch—that's actually doable. I can picture it working at a 50-person startup or within a product team at a larger company.

What I'm less sure about is whether the quality of the signal holds as the channel gets noisier. River works across 4,400 Slack channels at Shopify. At that scale, the question isn't whether public AI work creates learning—it's whether anyone can find and absorb the relevant threads when they need them. Jones suggests using AI to surface lessons from the channel itself, which is a nice recursion, but it also adds a layer of curation work that someone has to own.

The organizations that will actually pull this off aren't just the ones that flip the privacy default. They're the ones where at least one person with real judgment treats channel curation as part of their job, not a thing that happens automatically. That's a cultural bet, not just a tooling bet—and it's a harder sell than "River doesn't work in DMs."

Still: the core constraint is elegant. You can't always mandate culture change. You can sometimes mandate that the agent only works in public channels and let the culture change follow the friction. That's not naive optimism about good-faith participation—it's just product design applied to organizational behavior. And honestly? I find that more convincing than any number of AI adoption workshops.

The companies that figure out how to make individual learning compound at the team level are going to look very different from the ones where everyone got faster but nobody got smarter together. The gap is opening now, mostly invisibly, in private chat windows. River is one answer to it. The question is what your team's answer is.

Yuki Okonkwo is Buzzrag's AI & Machine Learning Correspondent. She covers the people and systems building the algorithmic future.