# Joveo Calls It Vibe Recruiting. Karpathy Meant Don't Read the Diff

_Joveo Tandem and Eightfold's AI Recruiting Teammate promise plain-English hiring agents. Here is the per-candidate evidence to demand before any outreach._

- Canonical URL: https://www.refolk.ai/blog/vibe-recruiting-joveo-tandem-missing-diff
- Published: 2026-11-18
- Reading time: 10 min
- Keywords: vibe recruiting, Joveo Tandem, AI recruiting agent, Eightfold AI Recruiting Teammate, AI sourcing tool evaluation

If you source for a living, typing the outreach email was never the hard part. The hard part is knowing whether you can check an AI agent's work before that email goes out.

On Nov 17, 2026, Joveo launched Tandem, an AI recruiting agent that takes a plain-English hiring goal and runs the workflow. CEO Kshitij Jain pitched it as Claude Code for recruiting and called the idea "vibe recruiting." That comparison only holds if recruiters get what developers get: a reviewable artifact behind every match, before any message is sent.

## What is vibe recruiting, and who is selling it?

Vibe recruiting is Joveo's name for typing a hiring goal in plain English and letting an AI recruiting agent run the hiring workflow. Three vendors announced some version of it within about a week in November 2026, and the category is clustering around high-volume hiring.

Jain made the analogy explicit: "AI tools like Claude Code transformed software development, turning natural language into production-ready code." The products themselves are quite different from each other:

| Product | Announced | Starting scope | Built around |
|---|---|---|---|
| Joveo Tandem | Nov 17, 2026 | Sourcing, scoring out of 100, email/SMS sequences | Job ads and volume hiring |
| Eightfold AI Recruiting Teammate | Nov 12, 2026 (preview, not GA) | Discovery inside the existing Eightfold Talent Network | Re-mining the employer's own database |
| Fountain Orchestra | Same week | Plain-language agents for hiring and scheduling | Frontline hiring |

The details that matter:

- **Tandem** scores candidates out of 100 and reaches out "with a personalized sequence via email, SMS, or both." It has a free plan, and setup "takes less than five minutes." It also promises that it "never rejects anyone," which protects inbound applicants but does nothing for a passive candidate who gets a sequence they never asked for.
- **Eightfold's Teammate** was announced on Nov 12, five days *before* Tandem, with CEO Joe Korngiebel quoted on the launch. It isn't generally available yet, so treat it as a preview.
- **Fountain Orchestra** lets companies build hiring agents by describing the work in plain language. Like Fountain itself, it's built for frontline roles.

The category is forming around volume work. Keep that in mind for the rest of this piece.

## Karpathy's vibe coding meant not reading the diff

Vibe coding, as Andrej Karpathy defined it on Feb 2, 2025, means accepting AI-generated code without reviewing it, and he limited it to throwaway projects. So when a recruiting product borrows the name, it's borrowing the mode where nobody reviews anything.

Karpathy's own words: "I 'Accept All' always, I don't read the diffs anymore." He called the approach "not too bad for throwaway weekend projects." In February 2026 he proposed "agentic engineering" as the name for the professional version of AI-assisted development and kept "vibe coding" for the throwaway kind.

Simon Willison put the line more sharply: "If an LLM wrote every line of your code, but you've reviewed, tested, and understood it all, that's not vibe coding in my book."

That changes the Claude Code comparison. Claude Code is valuable in a real codebase because every change it makes shows up as a diff a developer can read and check. The model writes the code, and the review is what makes it safe to ship. Drop the review and you have what Karpathy described: fine for a weekend toy, reckless for production.

> Karpathy kept vibe coding for throwaway projects. Your candidate market is not a throwaway project.

## A score out of 100 is not a diff

A diff shows the exact lines that changed, laid out so a reviewer can check each one. A score is a summary that asks you to trust the model. Recruiting agents are shipping scores and calling them evidence.

Both vendors use the word. Tandem "scores each one out of 100 with the evidence behind the score." Eightfold says the Teammate surfaces "the decisions that need human judgment, along with the evidence behind its recommendations." Neither says what that evidence actually contains.

Here's why that matters. You can't disprove a 78. A score folds a dozen signals into one number, and you can't tell which signal carried it: a keyword match on a job title, an inferred skill, or real work. A claim like "committed to tokio-rs/tokio in March" is different. You can disprove it in one click. That's the property a diff has and a score doesn't. Each claim can be checked on its own.

A developer who gets a pull request labeled "quality: 82/100" with no visible changes wouldn't merge it. Recruiters are being asked to send outreach on exactly that.

## Volume hiring and niche sourcing are different problems

When the pool has 126,486 people, a bad match costs almost nothing. When the pool has 196, one bad message burns half a percent of the market. Vibe recruiting is built for the first case and sold as if it covers the second.

Joveo's own explainer of vibe recruiting uses this scenario: "a TA leader who needs to hire 200 warehouse associates across three cities in 90 days." The workflow "begins with job advertising, optimizing ad placements across job boards and programmatic channels." That fits a company that "started out helping employers place and manage job ads." It's inbound volume work, and Tandem may well be very good at it.

Finding one specific passive engineer is a different job. Here is what the two ends of the market look like in Refolk's index of US professional profiles:

| Segment (US) | Count | Source |
|---|---|---|
| Warehouse Associate (current title) | 126,486 | Refolk's index |
| Technical Recruiter / Technical Sourcer | 21,671 | Refolk's index |
| Staff/Principal SWE with Rust listed as a skill | 196 | Refolk's index |
| Warehouse pool vs Rust staff pool | ~645x | 126,486 ÷ 196 |
| Technical recruiters per Rust staff/principal profile | ~111 | 21,671 ÷ 196 |

Accuracy matters differently at each end. In a warehouse search, a wrong match gets skipped and the ad keeps running. Speed wins. In the Rust search, every wrong message lands with someone already being chased by roughly 110 other recruiters, and you can't get that first impression back. Accuracy wins.

> **196** US Staff or Principal Software Engineers who list Rust as a skill
>
> In Refolk's index, against 126,486 US profiles titled Warehouse Associate.

The 196 is a floor. Many senior Rust engineers never tag the skill, so a tag search misses them, and that gap is the whole argument for open-web evidence. The people who do show up aren't sitting on job boards. In a top-25 sample of that pool, the most common employers are Meta (5) and Google (2), followed by NVIDIA, Figure, Shopify, Uniswap and PingCAP. These are heavily recruited, mostly passive people who don't answer job ads.

That leaves a founder or hiring manager who needs one of them in a bad spot. Job ads won't reach them. A score-and-sequence agent risks burning the few you can find. A contingency agency may spend weeks on the same thin list you already have and then charge a fee on top. Every week the seat sits empty, the team building around it slows down.

## What should a recruiting diff contain?

A recruiting diff is the specific, linked artifact behind each match: the commit, post, talk or profile line that explains why this person fits, which a recruiter can open and check in seconds. If the tool can't show it, you're vibe recruiting.

For every candidate, before any outreach, the tool should show:

1. **The artifact, linked.** A commit to a named repo, a merged PR, a conference talk, a blog post, or a specific profile line.
2. **The exact text that matched.** Quote it. "Led migration to Tokio" is evidence. "Strong Rust background" is the model's opinion.
3. **A date.** A 2019 commit and a commit from last month say different things about someone's current work.
4. **The requirement it satisfies.** Map each artifact to a line in your brief, the way a diff maps to a ticket.
5. **What's missing.** If nothing confirms seniority or location, say so. Showing what wasn't verified is the honest version of a test that didn't run.

That's the standard [Refolk](/) is built to. You describe the person in plain English, and each result comes back with the commit, post or profile line that put them on the list. Review becomes a quick scan of linked evidence instead of an afternoon of opening tabs.

> **Example search:** "Staff or principal engineers in the US who have shipped async Rust in production, with commits to tokio or a talk about migrating to it" - [run it on Refolk](https://www.refolk.ai/start?q=Staff%20or%20principal%20engineers%20in%20the%20US%20who%20have%20shipped%20async%20Rust%20in%20production%2C%20with%20commits%20to%20tokio%20or%20a%20talk%20about%20migrating%20to%20it)
>
> A ranked shortlist where every person comes with the linked commit, talk or profile line that explains the match.

## Where do the candidates come from?

Ask every vendor where its candidates come from before anything else, because the source limits what the agent can find. An agent that only searches your database or your applicant flow can't find the engineer who never applied.

Look at the starting scopes. Eightfold's Teammate "will initially focus on core recruiting activities, including candidate discovery within an organization's existing Eightfold Talent Network." That means it re-ranks people you already have. That's useful work: most ATS databases are full of good candidates nobody followed up with. It's still a re-ranking, not a search. Tandem's model, as Joveo describes it, starts with job ads and programmatic placement, so the funnel fills with whoever responds.

Neither approach reaches the Rust engineers missing from the 196 because they never tagged the skill. Those people show up through their work: commits, issues, talks, and the odd line in a profile summary. That's why Refolk searches the open web rather than an applicant pool: those engineers leave their evidence in their work, not in your ATS.

## The EU AI Act will make the evidence trail mandatory

Under the EU Digital Omnibus, the AI Act's high-risk rules for hiring systems now apply from Dec 2, 2027, pushed back from Aug 2, 2026. CV screening and candidate ranking are listed in Annex III, so per-candidate evidence trails stop being good practice and become a compliance requirement.

The delay gives buyers about a year of runway, not an exemption. Any AI sourcing tool you sign this quarter will probably still be in your stack when the rules apply. If it can only show a number per candidate, you'll be explaining that number to an auditor using evidence the vendor never showed you. Buy for the audit now and you won't need to migrate later.

## An AI sourcing tool evaluation checklist

Run every AI recruiting agent through the same six questions before it sends a single message on your behalf. The answers will tell you whether you're buying agentic sourcing or vibe recruiting.

1. **Where do candidates come from?** Your ATS, your ads, or the open web (commits, talks, posts, personal sites)?
2. **What is the evidence, exactly?** Ask for a sample candidate and click through. Is it a linked artifact or a score with a sentence of explanation?
3. **Is every claim dated?** Old evidence presented as current is the most common silent error.
4. **Can I approve before send?** Look for a review gate per candidate, not just a campaign-level "launch" button.
5. **How does it handle untagged skills?** Run a niche search (staff Rust, for example) and check whether it finds anyone beyond the obvious keyword matches.
6. **Can I export the evidence trail?** You'll need it for the hiring manager now and for the Annex III audit later.

Test with your hardest open role, not your easiest. Any tool looks good on a warehouse requisition. The 196-person search is the one that shows you what it can actually do.

## FAQ

### Is vibe recruiting the same as AI sourcing?

No. Vibe recruiting is Joveo's marketing term for prompt-driven, end-to-end hiring automation, and its own explainer centers on job advertising for high-volume roles. AI sourcing is the narrower job of finding specific people who fit a brief, often passive ones who never applied. The two overlap, but sourcing for niche roles depends on per-candidate evidence much more than volume hiring does.

### Is Eightfold's AI Recruiting Teammate available now?

Not generally. Eightfold announced it on Nov 12, 2026, five days before Joveo's Tandem launch, and it's not yet generally available, so treat it as a preview. Its initial scope is candidate discovery inside an organization's existing Eightfold Talent Network, which means it starts by re-ranking people already in your database.

### Can Joveo Tandem handle senior engineering searches?

Its public positioning points elsewhere: the explainer scenario of 200 warehouse associates and Joveo's background in job ads both point to volume hiring. Before trusting it with a passive-engineer search, run a niche query and check whether each score out of 100 comes with linked, dated artifacts you can verify yourself. If it doesn't, you're accepting the match on trust, which is the "Accept All" mode Karpathy kept for weekend projects.

---

_From the Refolk blog. More posts: https://www.refolk.ai/blog/index.md_
