Harness Engineering: Shifting from Human Oversight to Human-Guided AI Workflows

As organizations race to integrate AI agents into software development, the conversation is shifting from whether humans should oversee AI decisions to how humans can better guide them. Rather than keeping developers in direct control of every output, harness engineering proposes a model where humans stay involved by establishing clear quality standards and automated guardrails that allow AI to build more freely—but safely.

Kief Morris, a distinguished engineer at Thoughtworks, argues that the industry has become too focused on deployment speed while overlooking the engineering rigor needed to ensure AI-generated code is production-ready. "If you're using AI to build software, you need to make sure that you're building production-ready software," Morris said at a recent PlatformCon event in London, noting that most teams aren't incorporating fundamental best practices into their AI-assisted workflows.

The Problem With Losing Control#

The central challenge is that delegating code creation to AI agents can create distance between developers and what's actually being built. Without proper oversight mechanisms, teams risk losing visibility into code quality and system reliability. Morris emphasizes that this isn't about restricting AI, but rather about understanding what good engineering looks like before handing off work to agents.

"We've delegated so much of the code building to AI agents that we've become distant from the details of what they're building," Morris observes. This cognitive gap creates what researcher Margaret-Anne Storey calls cognitive debt—the hidden cost imposed on human teams who must understand and maintain AI-generated work without clear documentation or rationale.

Building a Continuous Delivery Framework for AI#

Morris proposes adapting time-tested CI/CD practices specifically for agentic workflows. The principle parallels traditional DevOps: when a bug surfaces in production, teams don't just patch the code—they trace back through the pipeline to understand why the issue slipped through and improve the system to prevent recurrence.

With AI agents, the same logic applies. If an agent produces inadequate code, the fix isn't simply to correct the output; it's to adjust the instructions, constraints, or quality checks so future iterations avoid the same mistake. This requires:

• Clear quality definitions that vary by team and business context