From MCP and Vibe Coding to Harness Engineering: How Did AI Native Engineering Evolve in One Year
Justy and Cody react to Birgitta Böckeler’s observation that AI-native engineering evolved from vibe coding to harness engineering in a year—shifting focus from prompt stitching to autonomous agents with built-in guardrails and risk assessment.
Script: Mistral Medium 3.5 128B Voice: Murf.AI Gen2
Transcript
Justy Okay so Birgitta’s saying we went from vibe coding to—what’d she call it—harness engineering in a year.
Cody Harness engineering. Yeah.
Justy And the jump is from pasting together Stack Overflow snippets to agents that can self-correct with… conventions and test suites?
Cody Right. Feed-forward tools like lazy-loaded skills save context window, feedback tools like static analysis catch mistakes before they ship.
Justy Mm-hm. So it’s not just speed—it’s confidence.
Justy Anyway. How was your flight from DC?
Cody Delayed. Two hours on the tarmac. Then the seat pocket was full of someone else’s trash. Classic.
Justy Ugh. And you still showed up to argue with me about AI?
Cody Someone’s gotta.
Justy But back to the evolution: monolithic MCP servers got replaced by lazy-loaded skills, CLI scripts, right?
Cody Yeah. And context window’s the bottleneck. So you chunk it into skills you pull in only when you need them.
Justy So that’s the feed-forward part?
Cody Exactly. Then feedback’s the tests and static analysis that tell the agent ‘hey you messed up row forty-seven.’
Justy And the UI split’s interesting too. Terminal tools like Claude Code for headless runs, Cursor for when you need the visual debug.
Cody I mean… that tracks. You don’t need a GUI if it’s just running scripts. But debugging a whole agent? Yeah. Good luck in a terminal.
Justy You say that like you’ve tried.
Cody I have. It’s NOT fun.
Justy So the real shift’s the harness. Conventions up front, tests on the back end. And the risk model—success odds, failure impact, detectability—feels almost… boringly practical.
Cody Boring’s good when you’re shipping code. But I dunno. ‘Harness engineering’ sounds like a buzzword we’ll all regret in six months.
Justy That’s such a Cody take.
Cody It’s ACCURATE.
Justy Fine. But who actually benefits? Teams that can afford the guardrails, or the ones already moving fast?
Cody The ones who can define their conventions AND their tests. If your repo’s a mess, the harness won’t save you.
Justy So it rewards discipline. Which… fair.
Justy Alright. Episode four seventy-five in the books. Try not to get stuck in a terminal this week.
Cody No promises.