Hey everyone, Neil here. You’re reading High-Signal Hiring. Hiring systems from 20+ years of global recruitment experience and 500+ technical hires. Zero noise and instantly actionable.
Last issue, we covered the FDE reject pool, the senior talent that the labs filter for you that nobody else is sourcing.
This week, a pattern I’m seeing in almost every screening I’m organising right now….
Every engineer I’m talking to says they use AI tools. Cursor, Claude Code, Copilot, the lot. Adoption is clearly no longer the question. 90% of developers use at least one AI tool in their workflow according to the Pragmatic Engineer’s 2026 tooling deep-dive, (Worth reading in full if you’re hiring engineers right now).
What they do with those tools is the question. And the split is wider than most founders realise.
Not a subscriber yet? Sign up here
| “Vibe Coding” vs “Agentic Engineering”
Andrej Karpathy coined the labels. “Vibe coding” is what he originally called it back in Feb 2025. You prompt the model, accept what comes back, and ship it if it runs. There’s no upfront spec or plan, and the diff never gets reviewed before merge. AI sits in the workflow like autocomplete with a chat box attached.
Then he walked it back in early 2026 and named the discipline that’s working. He called it “agentic engineering”. Same AI leverage, different posture. The engineer writes a short spec, plans the work, and orchestrates the agent against acceptance criteria. They review the diff line by line, catch what the agent got wrong, and re-run with corrections. AI is essentially treated as a junior engineer the senior is managing.
Both groups technically “use AI.” On output quality, however, they’re not in the same galaxy.
The cost of getting this wrong is huge. Individual productivity rises 21-55% with AI assistance. Organisational delivery stability is declining for teams without strong engineering foundations. Same tool, opposite effects. It makes a disciplined engineer measurably better and an undisciplined one measurably more dangerous. AI is a multiplier in both directions.
| Why you can’t see the split on a CV
Every CV in your inbox right now probably says “experienced with Claude Code, Cursor, GPT-5.” Every GitHub profile has an AI-built side project. JD filters like “must have AI coding tool experience” do nothing. With 90% adoption you’ve filtered out 10% of the market and learned nothing about the 90% you kept.
The filter has to happen live, in conversation, and it has to anchor in the candidate’s actual work, not just a hypothetical.
| Run the diagnostic in the founder screen
The mistake most founders make is saving this kind of test for the Deep technical interviews. By then you’ve already committed two hours of your team’s time per candidate. Run the diagnostic as early as possible. Inside the thirty-minute founder screen is ideal. The whole point of a ten-minute test is to filter at the cheapest point in the loop, before the expensive rounds.
One question does the work:
“Walk me through the most recent PR you shipped with AI agent assistance. From the moment you decided to build it through to the moment it merged.”
That’s it. Then you listen for three signals.
Signal one. Did they plan or did they prompt?
The agentic engineer opens with structure. “I wrote a short spec. I drafted three acceptance criteria. I broke it into two sub-tasks because one of them needed a different system prompt.”
The vibe coder opens with the prompt. “I opened Cursor and asked it to build the feature, then iterated from there.”
The difference is audible in the first sentence. One is describing a plan that existed before any code got generated. The other is describing a prompt and whatever came out.
Signal two. Did they verify or did they ship?
The agentic engineer talks about what they did to the agent’s output. They reviewed the diff line by line, wrote a failing test before generating, ran the result against edge cases, and caught the agent inventing a library call that doesn’t exist before it merged.
The vibe coder talks about the result. “I ran it and it worked.” Or worse, “I tested it locally and shipped it.”
Signal three. Can they describe a workflow or just a session?
The agentic engineer references infrastructure they’ve built around the work. They’ve set up a CLAUDE.md with house rules for the repo, saved custom slash commands for patterns they hit weekly, and wired in MCP servers for the tools their team uses. They talk about it the way an engineer talks about a build pipeline, because that’s effectively what it is.
The vibe coder describes one ad-hoc back-and-forth and has nothing systematic to point at.
Two strong signals out of three, you’re talking to an agentic engineer. Two weak signals, you’re talking to a vibe coder and the AI section of their CV is pure theatre.
A note on seniority. A staff engineer will answer this question with more depth and more named patterns than a mid-level. Calibrate to the level you’re hiring at. The signal is whether they’re describing a discipline at all, not whether they sound like a principal.
| What this isn’t
This isn’t a 60-minute paired session. This is the rapid front-end filter that saves you from ever getting to Depth interview (Main technical round) with the wrong candidate.
It’s also not a tooling trivia test. The infrastructure is a proxy for discipline. An engineer who’s never set up a CLAUDE.md but has a tight verification habit built around code review still passes. One who name-drops every tool in the stack but can’t describe a single workflow does not.
You’re testing for a habit, not a vocabulary.
| Make the swap this week
Open your current founder screen. Find the question that produces a CV recital. Replace it with the PR walkthrough. Watch what the next three candidates do with it.
You’ll feel the split inside the first ninety seconds. The plan-language candidate sounds different from the prompt-language candidate from the first sentence. Once you’ve heard the contrast twice, you can’t unhear it.
The hire that ships is the one with the discipline. The hire that breaks production at 2am on a public holiday is the one with the chat-box habit.
Cheers
Neil