For over two decades, enterprises have invested in observability tools to monitor software systems: dashboards, alerts, traces, and logs. Today the Site Reliability Engineering (SRE) teams responsible for supporting enterprise systems have plenty of data when issues occur. However, resolving them still requires significant manual effort to correlate those signals, find the root cause, and fix the problem.
Agentic AI has created a new generation of solutions promising to close that gap, which has significantly widened due to the complexity of modern software development.
The Reliability Crunch: More Complexity, Fewer Hands, Bigger Stakes
Fragmented Infrastructure: Kubernetes, microservices, multi-cloud, and hybrid environments have made modern software infrastructure vastly more complex and distributed. Troubleshooting issues requires engineers to navigate a web of system dependencies, data sources, and experts across functional teams.
Observability and Alert Fatigue: To maintain visibility over their large software footprints, enterprises often use multiple observability tools, each generating its own siloed telemetry and dashboards. When errors and incidents occur, SREs are overwhelmed with alerts, logs, and traces from each of these observability sources that they must analyze to resolve issues.
Talent Scarcity and Burnout: The accelerating pace of software development has created executive pressure to meet faster release schedules and underinvestment in system stability. As incidents occur more frequently, so does the overreliance on senior SREs with the required expertise and institutional knowledge to resolve them.
Amplified Impact of Incidents: Overwhelmed SRE teams often spend more time fighting downtime than improving system stability, which results in more frequent incidents, degraded system performance, and reductions in the overall engineering organization’s productivity.
Agentic AI SREs have emerged in direct response to these modern operational pressures, offering to reduce engineering toil, system outages, and incident resolution times.
Addressing the Shortcomings of Previous AI Solutions
Before the current boom of LLMs and agent-based solutions, the popularity of deep-learning neural networks led to a generation of “AI Ops” solutions that were used to automate IT operations using the large volumes of observability data that enterprises collected. AI Ops solutions and their underlying models were largely successful in making accurate predictions, but they suffered from three critical barriers to widespread adoption:
- Unique Incidents and Dynamic Systems: Faster software development led to constant changes in enterprise infrastructure and applications, so even the best predictive models required constant retraining to avoid degraded accuracy.
- Distributed and Heterogenous Telemetry Data: Training and maintaining predictive models became very labor-intensive due to the significant effort required to aggregate, prepare, curate, and label datasets.
- Black-Box Predictive Models: Previous AI models were successful in making predictions based on input data, but they lacked the explainability in critical decisions to build trust for adoption with infrastructure teams.
Large Language Models (LLMs) and the latest AI infrastructure have enabled more explainable and flexible solutions. The latest agentic SREs are not only better suited to consuming the unstructured data output from enterprise observability, but can dynamically extract enterprise context from code repositories, knowledge bases, and IT Service Management (ITSM) tools given the going improvements in context engineering infrastructure. SRE teams can focus on proactively improving system stability while AI agents react to and orchestrate the incident lifecycle – triaging alerts, detecting anomalies, correlating data, performing root cause analysis (RCA), proposing remediations, updating tickets, and even drafting pull requests.
