We’ve come a long way since Google’s Site Reliability Engineering book reframed uptime as an engineering discipline nearly a decade ago. Observability and automation have made building and running complex software systems saner and more reliable. What they haven’t done, though, is change the fundamentally reactive nature of troubleshooting production systems. AI agents may change that equation.

From reports to actions

Production systems should provide four basic services to ensure reliability and performance:

  • Detect a problem
  • Explain it
  • Help fix it
  • Learn and improve

Today’s observability solutions carry out the first two functions pretty well. Artificial intelligence can help us with the rest by creating what we call a “Vibe Loop.” The term borrows from “vibe coding,” a programming technique that takes an ad hoc approach to writing code. Large language models have turbo-charged vibe coding by letting developers issue natural language commands while machines do much of the grunt work.

Vibe Loop extends the same principles to observability. It’s a tight, AI-native feedback cycle between writing code, observing it in production, learning from it, and improving it quickly. Vibe Loop uses a network of agents to automate specific, repeatable tasks and suggest and even complete remedial actions. And it gets better over time.

In this new model of working with production systems, instrumentation is generated along with code to help human operators better understand a system’s behavior. AI surfaces and resolves basic spot problems automatically. Telemetry becomes adaptive as AI separates signal from noise. Postmortems are learning events that facilitate constant improvement. Engineers spend less time thrashing through logs and more time improving system performance.

There are three steps to implementing Vibe Loops.

Step 1: Prompt AI Code Generation Tools to Instrument

AI copilots and the OpenTelemetry standard for structured, vendor-agnostic instrumentation are changing how site reliability engineers (SREs) interact with observability tools. The combination lets them create prompts like:

  • “Write this handler and include OpenTelemetry spans for each major step.”
  • “Track retries and log external API status codes.”
  • “Count cache hits and DB fallbacks.”

Observability thus becomes a tool for detection and resolution, explaining its work as it goes along.

Step 2: Add Context

AI tools need more than raw telemetry to do their stuff. The open Model Context Protocol (MCP) developed by Anthropic is rapidly gaining traction as a standard and consistent way for applications to share information with AI models. It helps facilitate long or complex interactions by giving the model structured background information, acting as the glue between code, infrastructure, and observability.

SREs can use MCP to discover services, monitor for changes, identify the source of alerts, and search telemetry histories to learn how similar failures were previously handled. MCP gives AI the context to answer open-ended questions such as “Why is latency up?” or “Has this failure pattern happened before?” Answers can include reasoning about past incidents, correlated spans showing the full path of a request or workflow, and configuration changes. Engineers might spend hours piecing together the context for an event. Generative AI and MCP make the process instantaneous. AI agents can now gather context across multiple tools and reason about what they learn.

Step 3: Close the Vibe Loop

AI can not only help you better understand your production environments but also alert you to blind spots and offer corrective actions. It can note information you’re not capturing and offer to take that task off your hands or add missing attributes. It can even identify sources of inefficiency and offer to take corrective actions.

In the Vibe Loop, observability shifts from fighting fires and documenting actions to continuous discovery, diagnosis, and remediation. AI investigates incidents with the context of similar past events, surfaces potential root causes, proposes solutions, and helps SREs implement them on the spot. The quality of diagnosis and resolution improves with every incident. Engineering shifts from chasing traces to curating telemetry with intent. Developers can solve more of their problems without creating tickets and waiting in queues. Observability evolves from reactionary to adaptive.

Vibe Loop doesn’t replace engineers but empowers average engineers to operate at an expert level and experts to expand their scope and impact. For the first time, our tools are catching up to the complexity of the systems we run.