AI and SRE: Context Engineering as the New Pillar of Reliability

What happened?

In the Stack Overflow podcast episode aired on July 28, 2026, Ryan Donovan interviewed Asaf Savich, group manager of AI Engineering at Komodor. The conversation focused on how artificial intelligence is transforming site reliability engineering (SRE). The central theme was context engineering: the practice of providing AI models with the right information, at the right time, and with the necessary granularity to make reliable decisions in microservice environments.

Savich explained that as systems grow in complexity—with hundreds of interdependent services—human SREs become overwhelmed by data volume. AI can help, but only if it receives rich, structured context. Otherwise, models generate generic or incorrect responses that erode trust. According to Komodor data, 70% of Kubernetes incidents require crossing at least three different data sources (logs, metrics, change events), and teams spend up to 40% of their time manually gathering context. Context engineering aims to automate that collection and structure it so AI models can consume it effectively.

This approach is not entirely new: companies like Google and Netflix have internally developed context correlation systems for their SREs. However, the novelty lies in generalizing these practices through tools like Komodor, which integrate Kubernetes context with generative AI capabilities. Savich emphasized that the goal is for AI not only to detect anomalies but to understand causal relationships between changes, dependencies, and failures.

Why is it important?

The reliability of modern systems is critical for companies across all sectors. A service outage can cost millions and damage reputation. Traditionally, SREs relied on dashboards, alerts, and runbooks. However, with the adoption of microservices architectures and Kubernetes, the number of signals has multiplied. According to a CNCF study, 67% of organizations using Kubernetes experience incidents at least once a month, and the average resolution time exceeds two hours. Komodor's proposal and Stack Overflow's reflection point to AI being able to automate event correlation, anomaly detection, and even remediation execution, but it needs context that includes: change history, service dependencies, logs, metrics, and traces. Without deliberate context engineering, AI will be as blind as a human without data.

Savich compared this situation to the early days of observability: before tools like Datadog or New Relic, teams collected metrics manually. Now, context engineering is the next evolutionary step. Companies like Uber have already implemented machine learning systems to predict failures, but they face false positive problems due to incomplete context. Context engineering promises to reduce those false positives by ensuring models have all relevant information.

What consequences will it have?

The SRE role will evolve toward higher-value tasks: defining reliability strategies, designing context pipelines for AI models, and supervising autonomous agents that execute corrective actions. Savich described it as a shift from 'operator' to 'agent orchestrator'. This transformation is already seen at companies like Spotify, where SREs are adopting 'reliability as code' roles, and at Microsoft, which has integrated AI assistants into its Azure tools. Companies that invest in context engineering will be able to scale their operations without proportionally increasing their SRE team. Additionally, reducing false positives and faster resolution times will improve end-user experience. According to Gartner estimates, by 2028, 60% of organizations using AI for SRE will have reduced their critical incidents by 30%.

However, risks also arise: over-reliance on AI, biases in context data, and the need to maintain human oversight to avoid catastrophic automatic decisions. Transparency and auditability of models will be non-negotiable requirements. Savich warned that poorly implemented context engineering can generate 'contextual noise' that confuses models, leading to erroneous decisions. Therefore, he recommends starting with small use cases and continuously validating context quality.

What should readers know?

For SRE, DevOps, and platform professionals, this episode underscores the urgency of acquiring skills in data engineering, machine learning, and agent management. Tools like Komodor, which integrate Kubernetes context with AI capabilities, will become increasingly common. Also relevant is Stack Overflow's approach, which through its podcast and blog is positioning itself as a reference in the discussion about AI and software development. From TheVortiq, we recommend:

• Evaluate current data sources and their quality to feed AI models. Conduct a context audit: identify what data is available, its granularity, and freshness.
• Implement observability practices that capture rich context (events, changes, dependencies). Tools like OpenTelemetry can help standardize trace and metric collection.
• Train teams in prompt engineering and context flow design. Savich suggests that SREs learn to 'program' the context that AI receives, similar to how database queries are designed.
• Maintain a balance between automation and human oversight. Establish feedback loops where humans validate AI decisions and correct context when necessary.

As Savich noted: 'Context is the new oil for reliability.' The difference between a useful AI and a noisy one lies in the data we provide it. In an environment where 80% of telemetry data is never used (according to a Splunk report), context engineering becomes the key to unlocking AI's value in SRE.