Coinbase's One-Person Team: How to Source 30,400 Engineer-Designer-PMs
Brian Armstrong wants engineer-designer-PM hybrids running AI agents. Here's how to source the real ones using GitHub, LinkedIn, and side-project signals.
On May 5, 2026, Brian Armstrong cut 14% of Coinbase (about 700 people), capped the org at five management layers, told leaders to carry 15-plus reports, and announced the company would experiment with "one person teams" combining engineering, design, and PM, supported by fleets of AI agents. If you're a recruiter, your inbox now has a new req that reads like a unicorn order form. If you're a founder, you've probably been hiring this profile for two years already and watching most "full-stack product engineer" resumes wash out in week one.
The hard part isn't writing the JD. The hard part is that the cohort Armstrong is describing doesn't sit under a clean LinkedIn title, doesn't show up in a Boolean search for "Product Engineer OR Designer," and is increasingly hidden by the very tools (GitHub commit graphs, LinkedIn skill tags) that recruiters have leaned on since 2015.
What Armstrong actually asked for
Strip the marketing language and the Coinbase memo describes four operating constraints:
- No "pure managers." Every leader is a "player-coach" who still ships.
- Up to 15 direct reports per leader, flattening to five layers max.
- Engineers fluent in Copilot, Cursor, and Claude Code by the end of the week, not the end of the quarter.
- "AI-native pods" where one person directs agents across engineering, design, and PM scope.
That's a hiring rubric, not a vision statement. And it lines up with what's already happening elsewhere. Meta's new applied engineering team reportedly runs a 50-to-1 employee-to-manager ratio. Cloudwalk has engineers, PMs, designers, and operators all using Codex daily to turn specs into shipped code, scripts, fraud rules, or microservices in minutes. Life360's "AI-Native Backend" req explicitly demands a "heavy user of agentic workflows" who treats Claude Code as a first-class collaborator orchestrating agents to write specs, code, tests, and reviews.
The model is real. The question is whether your sourcing pipeline can find the people who actually live in it, or whether you'll spend six weeks reviewing LinkedIn profiles full of the word "AI" and end up no closer.
Why title search returns the wrong cohort
The instinct is to run a Boolean against LinkedIn: ("Product Engineer" OR "Founding Engineer" OR "Design Engineer") AND ("AI" OR "agents" OR "Cursor"). You'll get thousands of hits. Most are noise.
Our index shows roughly 30,400 people in the US holding the combined titles Founding Engineer, Product Engineer, or Design Engineer, heavily concentrated in NYC and the SF Bay Area. The interesting fact isn't the count. It's where they work. The top employers aren't FAANG. They're Mintlify, Thatch, Numeral, Athelas, Orbit, Known. Companies under 50 people that needed someone to ship a feature end-to-end because there was nobody else in the room.
This means the right sourcing query is closer to "engineers at companies under 50 employees who shipped a public artifact in the last 90 days" than to any title string. Title is downstream of company stage. Stage is downstream of behavior.
This is the kind of query that breaks LinkedIn Recruiter and X-Ray search but is trivial in plain English, which is why we built Refolk: describe the person ("US-based engineers at sub-50-person startups who've shipped a public side project using Claude Code or Cursor in 2025") and get a ranked shortlist instead of a 1,000-result cap of mostly-wrong profiles.
The signals that actually separate real hybrids from LinkedIn cosplay
If title is noise, what's signal? Five filters that survive contact with the cohort:
1. The side project has its own design system
Real hybrids don't ship bootstrap@3 demos. Look at their personal sites and side projects. Is there a coherent type scale? Do the empty states have illustrations? Did they bother with a 404? These take a designer five hours and an engineer five days. Someone who does both in a weekend is the cohort.
2. Solo-authored launches on Product Hunt or Show HN
A single name on the launch, a working demo video, copy that doesn't read like ChatGPT default voice, and a feature list that betrays product judgment (cut features, not added ones). This is the cleanest single signal we've found and it's not on LinkedIn at all.
3. Public commits inside the last 90 days, for managers
Armstrong's "player-coach" line is a literal sourcing filter. Pull every "Engineering Manager" or "Director of Engineering" in your target market and intersect with GitHub usernames showing commits to non-trivial repos in the last quarter. Most fail. The ones who pass are the managers worth poaching, because they're already operating the way Coinbase now requires.
4. Stars and forks on agent-orchestration repos
Addy Osmani's agent-skills repo is a useful lighthouse. So is anyone forking Claude Code configs, sharing Cursor rules, or contributing to MCP servers. This is a 2025-2026 signal that didn't exist two years ago and most ATS keyword filters don't catch it.
5. Title drift on the resume
Watch for someone whose last three roles drift Engineer → Product Engineer → Founding Engineer, or Designer → Design Engineer → Founder. The drift itself is the signal. People don't accidentally collect three adjacent disciplines. They optimize for it.
Title is downstream of company stage. Stage is downstream of behavior. Source on the behavior.
Why GitHub alone is getting worse, not better
A GitHub Next staff researcher put it bluntly: the PR and the issue are the wrong primitives that can't handle the speed, shape, or volume of agentic work, and very few people inside GitHub believe PRs and issues are ideal primitives for the future of engineering. Commit graphs are now inflated by agent output. A green wall of squares means less than it did in 2023.
The implication for sourcers: GitHub remains useful, but only when paired with off-platform artifacts (demo videos, launch posts, design portfolios). The combined picture is the signal. Neither half alone is enough anymore.
Stop screening on raw coding ability
Augment Code published their AI-native hiring rubric and openly noted that "raw coding ability as a standalone dimension" is missing. Their reasoning: as code gets cheaper to produce, the most expensive mistake is building the wrong thing. They evaluate four profiles instead, architectural-judgment IC, product-taste IC, model-depth IC, and learning-velocity early-career, and call those out as distinct hiring tracks.
If you're still leading with a LeetCode screen for AI-native engineer sourcing, you're filtering for the dimension that matters least and discarding candidates who are strongest on the dimensions that matter most. A founding engineer at Numeral isn't getting paid to invert a binary tree. They're getting paid to decide which three features to cut, write the migration plan, ship the UI, and tell the founder when the architecture is wrong.
Salesforce reports that AI natives are 4x more likely to use AI daily, deliver 3x faster than legacy managers, and drive a 40% increase in work quality. Treat vendor stats with appropriate skepticism, but the directional point holds: throughput and judgment, not raw coding speed, are what compound.
Where to actually pull from this quarter
The 700-person Coinbase pool is one source. It will be picked clean inside three weeks, like every public layoff list. The deeper pool is the 92,000-plus people that more than 100 tech companies have collectively cut in the first months of 2026, including AI-attributed reductions at Gemini Space Station, Pinterest, CrowdStrike, Chegg, Block, Klarna, and Citi.
The interesting subset inside that 92,000 is the senior ICs and player-coach EMs who were running 8-to-12-person teams before being absorbed into a flatter org or laid off entirely. They've already lived the megamanager shift (Gallup data shows average spans of control rose from 10.9 to 12.1 employees between 2024 and 2026), they've used Copilot in production, and they're available.
Specific places to mine:
- Life360's Foundry team and Meta's AI Native Software Engineer reqs as benchmarks for what the role looks like at scale, and as poach targets once people land there.
- Salesforce's Builder / Futureforce program, which is recruiting 1,000 graduates and interns into AI-native tracks. Early career, but a clean signal pool for the learning-velocity profile.
- Seed-stage shops in our index (Mintlify, Numeral, Thatch, Athelas, Known). Founding engineers at these companies have two-year tenures that are about to vest and they took the role specifically because they wanted the hybrid scope.
- GitHub Next and design-engineering Twitter, where people like Maggie Appleton (staff research engineer at GitHub Next, originally a designer) personify the hybrid. Her talk "One Developer, Two Dozen Agents, Zero Alignment" doubles as a candidate-graph anchor: who's citing it, who's responding, who's building in adjacent space.
For any of these, the workflow is the same: a plain-English description of the person, intersected across GitHub, LinkedIn, and the open web, then ranked. That's the exact loop Refolk runs, and it's why title-only tools cap out at 1,000 results of mostly the wrong people while the right 40 sit one query away.
The AI-washing caveat
Sam Altman noted at the India AI Summit that AI is reducing the need for employees, and also that companies are conveniently blaming AI when the layoffs aren't necessarily AI-driven. Coinbase is navigating a crypto downturn alongside its restructuring. "One person team" is partly an operating model and partly a narrative.
For sourcers, this matters in one specific way: don't assume every cut "AI-native" company is actually running AI-native pods. Some are running normal teams with a press release. Reference-check the workflow. Ask candidates what their last three weeks looked like. The ones who can describe an actual agent loop (spec → code → test → review, orchestrated, with specific tool names) are real. The ones who say "we use AI a lot" are not.
The OpenAI/METR finding that as of August 2025 leading models could complete 2 hours 17 minutes of continuous work at roughly 50% confidence, with task length doubling about every seven months, is the underlying enabler. It's also the thing your candidate should be able to talk about specifically, because if they're really running one-person teams, they're hitting the edges of that envelope every day.
The one-line summary
Source on company stage and public artifacts, not titles. Filter managers by recent commits. Discount LeetCode. Trust solo Product Hunt launches and side projects with their own design systems. Reference-check the actual agent workflow. The cohort is small (around 30,400 in the US under the most generous title definition, far fewer if you apply behavioral filters), but it's findable in a week if you stop searching the way you searched in 2022.
FAQ
Is "one-person team" actually viable or is it Coinbase marketing?
Both. The model is genuinely working at companies like Cloudwalk, where engineers, PMs, designers, and operators use Codex daily to ship scripts, fraud rules, and microservices in minutes. It's also being used as cover for cost cuts in companies that aren't actually running agent-orchestrated pods. For hiring purposes, treat it as a real operating mode at maybe a third of the companies announcing it, and reference-check the workflow before you trust the JD.
What's the single best non-LinkedIn signal for an engineer-designer-PM hybrid?
A solo-authored Product Hunt or Show HN launch in the last twelve months, with a working demo video, a designed (not templated) landing page, and a feature list that shows things were deliberately cut. That single artifact tells you the candidate can write code, make design decisions, and exercise product judgment without anyone holding their hand. Nothing else compresses three skills into one observable in five minutes.
How do I find player-coach EMs without 800 false positives?
Pull every "Engineering Manager" or "Director" title in your target geography, intersect with GitHub usernames, and filter for non-trivial commits to real repos (not docs or config) within the last 90 days. Most fail. The ones who pass are the cohort. This is precisely the kind of cross-source query that Boolean and LinkedIn Recruiter handle poorly and that a plain-English tool like Refolk handles in one prompt.
Is the Coinbase 700 worth sourcing directly?
For the first two weeks, yes, but it gets picked clean fast and the senior ICs are gone first. The bigger opportunity is the broader 92,000-person 2026 layoff pool, especially senior ICs and player-coach EMs from AI-attributed reductions at Pinterest, Block, Klarna, and others. They've already lived the flatter-org reality Armstrong is just now imposing.