Forget humans “in” the loop. Harness engineering puts humans “on” the loop.
The New Stack Jennifer Riggins ● Covered by 3 sources
Thoughtworks engineer Kief Morris says stop babysitting every AI-written line and start managing the pipeline instead. Skip that shift, and teams forget what "good" code even means.
Based on reporting by The New Stack, Jennifer Riggins — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Kief Morris has spent enough time around software delivery pipelines to know when something is quietly breaking underneath the hype. Speaking with The New Stack at PlatformCon in London last month, the Thoughtworks distinguished engineer argued that the industry's obsession with keeping humans "in the loop" of AI coding misses the point entirely. The real job now, he says, is putting humans "on" the loop — overseeing the system that builds the system, rather than eyeballing every commit an agent spits out.
Morris's worry isn't that AI writes bad code. It's that teams have gotten so far from the details of what agents produce that nobody's defining what good actually looks like anymore. He's blunt about it: organizations haven't been particularly good at articulating what "good" means even before AI entered the picture, and speed without that definition just multiplies the mess. His fix isn't some new framework bolted onto agentic workflows — it's the same continuous delivery discipline that's been around for more than a decade. If an agent produces something wrong, you don't patch the output; you fix the source and the pipeline so the same failure can't slip through again, the same way DevOps has always treated a production bug.
That idea connects to work from Margaret-Anne Storey, co-author of the SPACE framework, who recently described a Triple Debt Model to capture what AI-accelerated development is quietly costing teams. Technical debt lives in the code, cognitive debt piles up in the people trying to keep track of it all, and intent debt hides in the goals and design rationale nobody bothered to document. Morris frames the fear plainly: teams risk what he calls cognitive surrender, where the code runs ahead of anyone's understanding and becomes nearly impossible to safely change.
His proposed answer is something he calls harness engineering — building the surrounding structure that lets someone prompt an idea, let AI build it, then iterate, all while architectural decision records and clear guides keep the agents oriented. Underneath that sits operability: watching both lagging metrics like DORA data, user satisfaction and transaction costs, and leading metrics like failure rates, performance testing, and compliance audits. None of this is the flashy agentic story people want to tell right now, Morris admits, but skipping it just delays the moment you find out something's badly wrong, usually right when it's expensive to fix.
Morris's larger point is that the existing CI/CD pipeline, now 16 years into its life as an industry standard, is already the operability harness teams need — it just has to be treated as production-ready before AI agents get anywhere near it. He wants organizations to push agentic workflows hard, even further than feels comfortable, because that's how the real gaps surface. The resistance, he notes, shows up fast whenever someone suggests you might not need to look at the code at all.
My take — AI-written commentary, not fact-checked reporting
Kief Morris is basically saying the boring stuff was the important stuff all along, and he's right to say it loudly. Teams chasing agentic speed while skipping the pipeline discipline that's existed for over a decade aren't moving fast, they're just deferring the bill. The idea of cognitive surrender should scare people more than it does — losing track of your own systems isn't innovation, it's just a slower-motion outage waiting to happen.
Read more about this at: The New Stack