AI-assisted development is a tantalizing opportunity for any software organization. While it clearly offers speed, it raises an immediate technical debt question: how do we prevent faster delivery from quietly degrading maintainability?

This talk shares the experience of loveholidays, one of Europe’s largest online travel companies, operating at scale with multiple autonomous teams delivering continuously to production. Between June 2025 and March 2026, loveholidays scaled the adoption of agentic coding, primarily using Claude Code. During this period, the rate of AI-assisted commits increased from 0% to 45%, with a 20% increase in velocity, as measured by year-over-year deployments to production.

To ensure that AI-generated code remained maintainable, we collaborated with CodeScene, a software intelligence company specializing in socio-technical analysis and maintainability assessment. CodeScene’s CodeHealth metric was introduced as explicit quality guardrails inside the agentic loop. Initially, these guardrails were enforced via the CodeScene CLI. In January 2026, we migrated to an MCP-based integration, providing a more robust and developer-friendly feedback channel for coding agents.

The agentic feedback loop evolved organically. Expert developers responded to feedback on experiments in their approach to agentic coding and trialled a variety of prompting techniques, culminating in tooling declarations for the CodeScene CLI and the claude.md file. Consistent signals indicate that the current generation of LLMs is well trained to leverage CodeHealth feedback, leading to a virtuous coding cycle.

The agents iterated on any low-CodeHealth code additions until they reached a predefined threshold, intentionally set at a high value (a perfect 10). While overall code production increased, our results show that CodeHealth scores increased as well. During the talk, we will share quantitative results from CodeScene, complemented with DORA metrics and qualitative insights from developers.

The key takeaway is simple: successful agentic coding requires responsible engineers and measurable quality guardrails. This talk shows how loveholidays and CodeScene made agentic development work in practice - without trading long-term maintainability for short-term speed.