Operations | Monitoring | ITSM | DevOps | Cloud

AI coding assistants like Gemini are changing how developers write code. They can generate entire functions, debug tricky issues, and help you move faster than ever before. But with that speed comes a new challenge: how do you make sure AI-generated code actually works? AI assistants are powerful, but they’re not perfect. They can introduce subtle bugs, miss edge cases, or generate code that breaks existing functionality. That’s where CI (continuous integration) comes in.

AI-powered coding tools are changing how developers work. Tools like Claude Code can write functions, refactor code, and build features through natural conversation, often faster than you could type them yourself. But speed creates its own risks. AI-generated code can contain subtle bugs, reference packages that don’t exist, or misuse APIs in ways that only surface at runtime. That’s where continuous integration comes in. CI is a safety net that lets you move fast confidently.

In the world of enterprise software, the transition from traditional DevOps to modern AI-driven delivery is less like a flip of a switch and more like a high-stakes kitchen. As Devan Shah, Chief Architect at IBM, puts it: the ingredients have changed from food to code, but the need for a precise, governed process remains the same.

This article is part of our series on how Datadog’s engineering teams use LLM Observability to build, monitor, and improve AI-powered systems. Organizations are building AI agents that help users automate work, analyze data, and interact with complex systems through natural language. As these agents become more capable, they also become more complex and exposed to risks such as prompt injection, data leaks, and unsafe code execution.

The production environment has become a minefield of code nobody really understands. Here’s what’s happening: Development teams are using Claude Code, Cursor, and GitHub Copilot to ship features at 10x their previous velocity. Product managers are ecstatic. Business stakeholders are thrilled. And somewhere in a war room at 2:17 AM, an SRE is staring at a stack trace for code that was AI-generated three weeks ago, trying to figure out why the payment service just fell over.

Model Context Protocol (MCP) servers act as adapter layers between clients and AI based workloads. MCP installation into an IDE, such as Cursor, brings a wealth of information directly into the developers primary tool, minimizing context switching and, especially in the world of observability, bringing telemetry closer to the code. MCP is not without its limits. These limits initially seem trivial, but in time, some of the inherent limitations to a basic MCP implementation become apparent.

In today’s highly interconnected systems, the subtle relationships between services are rarely obvious. Modern, complex architectures generate telemetry that functions less as “flashing signs” and more as faint “breadcrumbs” to be followed across a vast network of signals. In 2025, about two-thirds of outages involved third-party systems like cloud platforms and APIs.