The 720:1 Gap: Why Boolean Sourcing Can't Find AI-Native Engineers

Scroll the July 2026 "Who is hiring?" thread on Hacker News and you will hit the same phrase every third post: "AI-native," "ships with Claude Code," "uses AI to generate 90%+ of code." The JDs have raced ahead of the resumes, and every sourcer who leans on LinkedIn Boolean is now searching a vocabulary that ~88% of qualified engineers do not use about themselves.

That is the arbitrage. The engineers exist. The words to find them do not.

The gap in one number: 91% use it, 12% say it

CodeSignal's March 2026 survey of 450 U.S. software engineers found 91% already use agentic AI coding tools at work and 75% have shipped production code partially or primarily generated with AI in the last six months. A 2026 resume audit cited alongside the CodeSignal launch found fewer than 12% of engineers mention agentic AI tools on their resumes.

That is a ~7.6x delta between what engineers do and what their profile says they do. Every keyword string you write on LinkedIn Recruiter is filtering on the wrong end of that ratio.

7.6x

The mechanism is boring and predictable. Resumes are lagging documents. Engineers update them when they change jobs, not when they change workflows. Claude Code shipped its 1M-token Max tier in the last year. Cursor's rules files are newer than most of the resumes in your ATS. The tools moved faster than the paper.

What the July 2026 HN thread is actually asking for

The new-style JD is not asking for "AI experience." It is asking for daily operator fluency with specific agentic tools, and it treats that fluency as a hard filter, not a nice-to-have.

A few live examples from the July 2026 thread and adjacent 2026 postings:

• Namecoach / Euphonia founding voice-AI role: "You should be comfortable shipping with agentic dev tools like Claude Code and Cursor."
• Anthropic Product Designer for Claude Code: candidates must be "AI-native in how you work," already using Claude Code or similar tools to extend what they can build.
• Cursor, Vercel, Linear, Lovable, Perplexity, and OpenAI 2026 postings recruiting "Super ICs" who span product strategy, design, and engineering with agentic tools as the default surface.

The pattern is the same across all of them: the tool name is in the requirements block, not the "bonus" block. And the reason it is a hard filter is downstream of team math, which we will get to.

Refolk's index says the Boolean pool is 12 people

Type the exact JD language into a Boolean search and the pool collapses to a rounding error. In Refolk's index of U.S. professional profiles, only 12 people combine "Cursor," "Claude Code," and "AI-native" in their current headline or summary.

Twelve. In the entire United States. Split across "AI-Native Product Engineer" (1), Product Designer (2), Software Engineering Lead (1), and a scatter of individual contributors.

Widen the filter and the numbers get more useful, but still nowhere near the 91% survey number:

GitHub Copilot is still roughly 3.4x more discoverable on profiles than Claude Code and 6.7x more than Cursor, even though Claude Code is the tool most July 2026 JDs are converging on. That is the resume-lag problem in a single row: engineers listed Copilot two years ago and have not touched the skills section since.

The derived stat that matters: the ratio between Copilot-tagged profiles and the full "AI-native operator" intersection is roughly 720:1. If you filter for the words on the JD, you throw out 719 out of every 720 candidates who could actually do the job.

Why founders are writing "90% AI-generated" into the requirements

The aggressive JD language is not a filtering strategy. It is a headcount strategy. AI-native startups can afford exactly one kind of hire, and it is not the person who needs a ramp.

Ravio's 2026 analysis of AI-native startups at Series A and B found these companies maintain a median of 73 employees versus 98 for non-AI peers, making them 34% leaner. Pragmatic Engineer documented a Series C startup that restructured a 12-person engineering team into 3 people using Cursor and Claude, hitting a 40% increase in development velocity despite the 75% headcount cut.

Every hire on a lean AI-native team carries roughly 1.5x the load. Founders are not filtering for cultural fit. They are filtering for output density.

That is why the language got sharp. When each engineer is doing the work of one and a half, you cannot afford a non-native ramp on the first week. You need someone who already ships with Claude Code the way a 2018 hire already shipped with Git.

Boolean is broken because the vocabulary is new

Boolean sourcing works when the requirement is a stable noun. "Python" has meant Python since 1991. "AI-native" has meant something specific for maybe fourteen months. That is the entire problem.

Three follow-on effects sourcers should internalize:

1. Keyword strings over-index on hobbyists. The engineers who bothered to update their LinkedIn headline to "AI-Native Product Engineer" are disproportionately the ones with time to update their LinkedIn headline. Senior operators shipping five PRs a day are not among them.
2. Tool-loyalty filters miss the senior bar. Most AI-native senior engineers run both Claude Code and Cursor, mapped to the right surface. Hard-filtering on one collapses the pool by ~50% and specifically excludes the hybrid-fluent operators most JDs actually want.
3. The intersection filter is empirically capped. Refolk's 12-person "Cursor + Claude Code + AI-native" set is the empirical ceiling of the profile-text approach. You cannot Boolean your way past it.

This is the exact gap Refolk closes for AI-native engineer hiring. Instead of stringing together ("Claude Code" OR "Cursor") AND "AI-native" AND senior, you describe the person in plain English ("senior backend engineer who ships daily with agentic tooling, has open-source receipts, and has worked on a lean team under 15") and get a ranked shortlist pulled from GitHub, LinkedIn, and the open web at once.

The receipt beats the skill tag

The most reliable "is this person AI-native?" signal is not on the resume at all. It is in their GitHub. The files an AI-native engineer commits are load-bearing artefacts, and they are trivial to search for if you are looking in the right place.

Signals that beat any LinkedIn keyword:

• CLAUDE.md in a repo root: Claude Code project instructions. If the engineer wrote one that is more than 30 lines, they are running Claude Code as their primary loop.
• .cursorrules or .cursor/rules/: Cursor project rules. Same signal, different tool.
• AGENTS.md: emerging convention for multi-agent project context, common in Aider and Codex workflows.
• MCP config files (mcp.json, server configs): the engineer has wired Claude to their own tooling, which is a two-tier-above-average signal.
• Commit cadence patterns: bursts of 10+ commits per hour on a solo repo, especially with terse messages, correlate with agentic loops.

Not knowing these files exist is the AI-native equivalent of not knowing what a .gitignore is. Which is why Claude Code Cursor recruiting works so much better when the source of truth is the code, not the profile.

Refolk indexes these artefacts across GitHub and the open web, which is the mechanism that closes the 720:1 gap. A sourcer asking "engineers with CLAUDE.md in a repo pushed in the last 30 days, based in the US, 5+ years experience" gets a list that Boolean cannot produce, because none of those signals are on LinkedIn to begin with.

Assessments are eating the resume

When 75% of engineers ship AI-generated code and ~12% say so on paper, the only reliable way to distinguish real fluency is to watch someone work. That is why the assessment layer of the funnel is growing so fast.

CodeSignal launched agentic coding assessments on April 2, 2026. In the six months prior, tens of thousands of candidates completed AI-assisted coding assessments on the platform, and roughly one-third of CodeSignal customers adopted an AI-assisted format during 2025. CoderPad's 2026 State of Tech Hiring report, based on more than 650 respondents, found technical assessments up 48% globally versus mid-2023 while U.S. technical hiring activity is up 90%.

48%

The philosophical shift matters. Traditional assessments graded the output. Agentic coding assessments grade the interaction with AI: what the candidate prompted, how they reviewed, when they overrode, whether they caught the model's mistakes. Tigran Sloyan, CodeSignal's CEO, framed it plainly: "Engineers are no longer coding alone; they're working with AI agents, and the best ones know how to get the most out of them."

That is the read-through for sourcers. The resume is now a coin-flip. The receipt (GitHub artefacts) narrows the pool. The assessment (a 45-minute recorded prompt-and-review session) confirms it. Boolean does none of the three.

A working playbook for sourcing AI-assisted developers

If you take one thing from this piece, take the order of operations. Sourcing AI-native engineers is a four-step funnel, and none of the steps is a LinkedIn keyword string.

1. Start from receipts, not resumes. Search GitHub and the open web for CLAUDE.md, .cursorrules, AGENTS.md, and MCP configs in repos active in the last 60 to 90 days.
2. Cross-reference to identity. Match the GitHub handle to a LinkedIn or personal site to confirm location, seniority, and current employer. This is where most Boolean-first pipelines fall apart, and it is where Refolk's cross-source lookup earns its keep.
3. Ignore tool-loyalty in the JD. Ask for "agentic tool fluency" and let the candidate map to the right surface (Claude Code for terminal-heavy work, Cursor for IDE-heavy work, Windsurf, Aider, or Codex CLI where they fit). The 12-person intersection filter is the ceiling of doing otherwise.
4. Use an agentic assessment as the final filter. A recorded 30 to 60 minute session with the candidate driving Claude Code or Cursor against a real problem tells you more than any keyword match, and it is the only step that catches the ~79 point gap between what engineers do and what their profile says.

Sourcing communities also matter here. The Cursor community, Claude Code Discord, and Windsurf forums contain engineers actively pushing the boundaries of AI-assisted development. None of them are searchable on LinkedIn. Most of them do not have "AI-native" anywhere in their profile. All of them have receipts.

FAQ

Why do so few engineers list Claude Code or Cursor on their LinkedIn?

Resumes and LinkedIn profiles are lagging documents that most engineers update at job-change time, not at workflow-change time. Claude Code and Cursor became the default agentic tools inside the last twelve to eighteen months, which is shorter than the median gap between profile edits. Combine that with the fact that agentic tooling still reads as "cheating" to some hiring managers, and you get an active disincentive to list it. The 91% vs. 12% gap is a discoverability problem, not a competence problem.

What Boolean string actually works for finding AI-native engineers?

None of them work well on LinkedIn alone. The best keyword-only approach is to look for adjacent signals ("solo founder," "prototyped," "shipped MVP in") plus recent GitHub activity, but even that is a poor filter. The reliable approach is to search code artefacts (CLAUDE.md, .cursorrules, MCP configs) on GitHub and the open web, then cross-reference to profiles. This is why plain-English search across sources (the model Refolk uses) closes the 720:1 gap that Boolean cannot.

Should our JD require Claude Code specifically, or Cursor, or both?

Neither, as a hard filter. Refolk's index shows 2,516 U.S. profiles with Claude Code and 1,280 with Cursor, and the intersection with "AI-native" collapses to 12. Requiring a specific tool by name filters out most senior operators, who fluidly switch between tools by task. The better JD language is "daily fluency with agentic coding tools" and then a portfolio or assessment step where the candidate demonstrates it on their preferred stack.

How do agentic coding assessments differ from traditional coding assessments?

Traditional assessments grade the final code. Agentic coding assessments grade the interaction: the prompts the candidate wrote, the outputs they accepted or rejected, the moments they overrode the model, and the review discipline they showed. CodeSignal, which launched agentic assessments on April 2, 2026, gives hiring teams a transcript of the candidate's session with the AI. Given that 75% of engineers already ship AI-generated production code but only about 12% say so on paper, a recorded agentic session is currently the only reliable way to distinguish real fluency from resume noise.