Stop correcting AI code. Build the system agents need.
The New Stack Jennifer Riggins ● Covered by 3 sources
Devs are told to quit babysitting AI code line by line. Instead, fix the system feeding your AI, not the prompt you typed.
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
Patrick Debois, the guy who literally coined the term DevOps back in 2009, is now making a similar case for AI coding tools. His pitch, delivered at PlatformCon London this summer, is blunt: when an AI agent botches a task, don't just rewrite the prompt or patch the code by hand. Fix the system that's feeding it context in the first place.
That sounds like a small mindset shift, but Debois frames it as a full lifecycle change. The old software development lifecycle, he argues, is becoming a context development lifecycle — a set of nested loops covering how context gets generated, tested, packaged, and observed. Where CI/CD used to be one feedback loop, now you've got layers of them, all built around the idea that AI agents run on probabilistic systems, not deterministic ones, so the old rules about fixing bugs one at a time stop working.
The scaling problem is where this gets organizational rather than personal. Debois points out that different teams feed agents wildly different inputs — codebases, post-mortems, Slack threads, architectural decision records, compliance docs — and none of it is coordinated. His advice: stop optimizing context for your own laptop and start asking whether it works for the whole team, then the whole platform. He compares it to how DevOps itself spread, starting with one enthusiast, growing into a team habit, then eventually getting baked into the platform everyone uses.
He's also skeptical of the industry's default fix for shaky AI output, which is basically throwing more AI at the problem. Asking one LLM to grade another LLM's code, he says, is a shallow check — it's just reading the code, not running it. The deeper fix is what he calls harness engineering: wrapping agents in logs, metrics, and traces so they can see the actual consequences of what they wrote and self-correct, rather than needing constant human correction.
Done right, Debois says, these harnesses stop being a single team's private tool and become shared infrastructure — a common way of testing and verifying that other teams start building on. Most companies will end up with multiple overlapping pipelines rather than one unified system, but the point stands: building the scaffolding AI agents need turns into a forcing function for better collaboration across the whole engineering org, not just a productivity hack for individual coders.
My take — AI-written commentary, not fact-checked reporting
This is the DevOps playbook with a new coat of paint, and that's exactly why it'll work for the orgs that actually do it and go nowhere for everyone who treats it as a Slack thread and a Notion doc nobody updates. The uncomfortable truth here is that AI coding tools expose which engineering orgs already had a mess of undocumented tribal knowledge, and now that mess is what your agents are inheriting too.
Read more about this at: The New Stack