Coinbase's Mux Has 197 Power Users. That's Your 2026 Shortlist.
Coinbase's Mux power users ship 3.5x more PRs than baseline. Here's how to find their equivalents at Stripe, Block, and every AI-native company.
Coinbase's engineering blog quietly buried the most important sourcing signal of 2026 inside a post about a side project. Mux, an internal multi-agent orchestration tool that started when one engineer scratched an itch, now has 600+ users across every org, 197 of them classified as power users, and those power users are merging 3.5x more PRs than baseline. If you're still typing "AI Engineer" into LinkedIn, you are searching for cosplayers while the real shortlist ships code.
The 197 that matter
Read the Mux numbers slowly. In April 2026 alone, the tool powered 600+ users across engineering, PM, and design, with 335 active and 197 power users driving 5,068 merged PRs across 461 repositories and 10 orgs. Power users merged 39.6 PRs versus a baseline of 11.4. That is not a productivity story. That is a sorting function.
The 600 is the user base. The 197 is the shortlist. The 403 in between tried Mux, got some value, and moved on. The 197 changed how they work. That's the population you want to hire, and none of them have "AI Engineer" in their headline.
The Coinbase blog is careful about attribution, but the sourcing pattern is obvious. Mux started as one engineer's side project, got shared in a Slack channel, and grew organically. There was no adoption campaign. That means the 197 power users are traceable in commit history, in internal-tools plugin contributions, and in the "many users started contributing back" cohort the blog names but doesn't list. Those contributions leak to public GitHub via adjacent infra work, conference talks, and open-source forks of the same primitives.
Why "AI Engineer" is a negative signal now
Filter Refolk's index for US-based Senior or Staff engineers who actually list Claude Code, LangGraph, or Cursor as skills and you get roughly 3,026 candidates. The dominant title is Senior Software Engineer. Not AI Engineer. Not ML Engineer. The people building agent fleets at Coinbase, Stripe, and Block did not rebrand. They kept shipping under the title they had, because the title was never the point.
This is the trap in AI-native engineer sourcing. Recruiters index on headlines because headlines are searchable. But the taxonomy on LinkedIn is a lagging indicator that hasn't caught up to how work actually gets done in 2026. Coinbase's own framing calls these engineers "orchestrators of agent fleets," which is not a title, is not a job family, and is not something anyone writes in a bio.
The 600 is the user base. The 197 is the shortlist. Sourcing by title finds neither.
If you want to find them, you have to search behaviors: which internal tools they helped extend, which agent harnesses they've committed to publicly, which developer-tools Slack communities they show up in, which Lenny episodes they've been quoted in. That is exactly the plain-English query problem Refolk was built for. You describe the person you want ("Staff-level backend engineer, has publicly contributed to Goose, Claude Code plugins, or LangGraph, worked at a company with an internal coding agent") and get a ranked shortlist instead of a title match.
The Goose to Minions to Mux lineage
The best proof that these engineers exist as a coherent population is the lineage of the tools themselves. Block open-sourced Goose. Stripe forked Goose into Minions, stripped out the interruptibility and confirmation dialogs, and optimized for one-shot task completion. Minions now ships 1,300+ PRs a week with zero human-written code across hundreds of millions of lines of mostly Ruby. Coinbase's Mux sits in the same family, alongside Shopify's Roast (open source), Google's internal Agent Smith (access had to be restricted because it was too popular), Meta's REA, OpenAI's Harness, Ramp's Inspect, Browserbase's bb, and Coinbase's earlier Forge, which hit 5% of merged PRs and cut PR cycle time from 150 hours to 15 hours.
That is 15 confirmed in-house agent systems, no longer experimental. Every one of those companies has somewhere between 50 and 300 internal power users of a Mux-equivalent. Do the arithmetic and you get a total addressable shortlist of roughly 5,000 people hiding in plain sight across ten companies you already know how to find.
The contributor list beats the user list
Here is the non-obvious move. Do not chase Mux users. Chase Mux contributors. The Coinbase blog is explicit that many users started contributing back with integrations to internal infrastructure, plugins for team-specific workflows, and bug fixes. Contributors are a strict subset of power users. They are the ones who not only used the tool 3.5x more effectively but also had opinions about how to make it better and the seniority to act on them.
Contributors surface in three places you can actually search. First, adjacent public repos where the same engineers push infra work. Second, developer-tools conferences and meetups where they present internal case studies. Third, the podcast circuit, where Coinbase's Chintan Turakhia is already publicly explaining "the speed run technique that got 100 engineers to push 70 PRs in 15 minutes" and how to identify power-user behaviors. Every one of those signals is a name.
The Coinbase exit window is open and mispriced
Armstrong is not subtle about the plan. He's eliminating pure managers, replacing them with player-coaches, and building AI-native pods (some staffed by a single person) that deploy AI tools across engineering, design, PM, compliance, and support. Each manager will oversee 15+ direct reports. Coinbase claims AI now assists in 40% of code written and cut compliance false positives by 90%. Every engineer got Cursor, Copilot, OpenCode, and Claude Code by early 2026. Armstrong asked engineers to onboard by end of week. Some who missed the deadline "got fired."
Read that last sentence twice. The cohort that survived is, by definition, the cohort that adopted. And the cohort that thrived (the 197 Mux power users, the Forge contributors, the ones whose PR throughput jumped) is the exact population that will now be pulled into pods so small that a chunk of them will leave rather than run compliance, design, and eng out of one seat.
Kore1's field data on the aftermath is the sourcing tell. Three buyers passed in the same week on the same senior Coinbase engineer: IC4, 8 years in, Onchain Payments, $487K total comp. His profile read "too Onchain" and his screenshot looked expensive. The day after the third pass he signed with a Series C stablecoin startup in NY in five days, below Coinbase base but above his RSU on the back end. That is a mispricing you can arbitrage right now if your reqs are Staff-level and you move before the profile refreshes.
The lesson is not "hire from Coinbase." The lesson is that multi-agent developer hiring is a game about spotting adoption curves before comp data catches up. The people who onboarded Mux in month one, who wrote the first plugins, who spoke about it internally, are the same archetype as the people at Stripe who built Minions and at Block who maintain Goose. They move in a small, dense graph that Refolk lets you traverse in one query instead of ten Boolean strings.
What to actually search for
Forget "AI Engineer." Here is the behavior stack that identifies real agent fleet orchestration engineers in 2026:
- Public contributions to Goose, LangGraph, Claude Code plugins, Cursor extensions, or Aider.
- Employment at one of the 15 companies with an in-house agent harness (Coinbase, Stripe, Block, Shopify, Google, Meta, OpenAI, Spotify, Cloudflare, Uber, Ramp, Browserbase, Anthropic-adjacent shops, and a handful of scaleups).
- Titled Senior Software Engineer or Staff Software Engineer, not "AI Engineer." Rebranders are the false positive.
- Conference or podcast appearances discussing internal tooling, agent evaluation, or developer velocity metrics. Chintan Turakhia's Lenny episode is the template.
- Commit history that shows a jump in PR count in mid-to-late 2025 without a corresponding job change. That's the Mux effect showing up in the graph.
- Side-project agents on personal GitHub. The Mux origin story is not unique. Every real power user has a scratch-your-own-itch repo.
Steve Kaliski at Stripe is the public archetype. Six and a half years on developer tools and payment infrastructure, part of the team that built Minions, publicly discoverable, not titled "AI Engineer." Every company on the list has three to ten Steves. Your job is to name them before someone else does.
That's the whole thesis of how to source AI power users right now: stop searching for the label and start searching for the behavior. Refolk exists because the label-based tools cannot do this. Ask in plain English for "Staff engineers who contributed to internal coding agents at Coinbase, Stripe, or Block and are currently in market," and the ranked list is your shortlist, not a title-match wall of noise.
FAQ
Why is "AI Engineer" a bad search filter in 2026?
Because the highest-leverage engineers building agent fleets at Coinbase, Stripe, and Block never rebranded. Refolk's index shows the dominant title among people who actually list LangGraph, Claude Code, or Cursor as skills is Senior Software Engineer. The people who changed their headline to "AI Engineer" are disproportionately mid-level ICs chasing the trend, not the Staff+ ICs who own internal tooling. Filter by tool and seniority, not by title.
How do I find Mux "power users" if Coinbase doesn't publish names?
Triangulate. The 197 power users overlap heavily with contributors to Mux itself, and contributions leak into public GitHub through adjacent infra work, plugin repos, and conference material. Cross-reference Coinbase engineers who appear on podcasts like Lenny's (Chintan Turakhia is one), who speak at developer-tools meetups, or whose commit graphs jumped in mid-to-late 2025. That funnel of behavioral signals is exactly what a plain-English query in Refolk resolves to a ranked list.
Are the Coinbase layoffs actually a hiring opportunity?
For Staff-and-above reqs, yes, and it's currently mispriced. Kore1 documented buyers passing on senior Coinbase ICs because comp screenshots looked expensive and profiles read "too Onchain," while Series C startups closed the same candidates in five days below base. If your req is below Staff, you're competing for a role Coinbase's AI-native pod restructure just proved you don't need. The war is over the top 5%.
What's the total addressable shortlist across all the in-house agent tools?
Roughly 5,000 engineers. There are 15 confirmed in-house agent harnesses across Coinbase, Stripe, Block, Shopify, Google, Meta, OpenAI, Spotify, Cloudflare, Uber, Ramp, Browserbase, and a few others. Each has 50 to 300 internal power users of a Mux-equivalent. The contributor tier (the people who extend the tools, not just use them) is closer to 1,500. That's the real 2026 sourcing pool for multi-agent developer hiring, and almost none of them are searchable by title.