Recent years have seen generative AI (GenAI) models integrated into monitoring, observability, and ITSM solutions to assist IT professionals with text generation, remediation suggestions, insights into system health, and more. The next step in the evolution of AI involves architectures that see multiple AI agents working together to solve problems. Agentic AI promises to push artificial intelligence from an assistive tool to an increasingly autonomous role in IT operations, reshaping how enterprises manage their environments.
The Rise of Agentic AI
So, how is Agentic AI different? Well, traditional large language models (LLMs) primarily focus on generating text, answering queries, and offering suggestions. They respond directly to user queries in a single step, often without iterative reasoning. In contrast, Agentic AIs employ strategies like chain-of-thought reasoning to break down tasks, evaluate intermediate outcomes, and adjust actions, leading to more refined solutions for complex workflows. AI agents are powered by LLMs but extend their capabilities through integration with other tools and data. Far from being standalone systems, Agentic AI sees LLMs evolving into goal-oriented collaborators that can understand objectives, adapt to context, and autonomously navigate workflows.
Moving Beyond Individual Agents to Compound AI Systems
Compound AI systems, where multiple specialized agents collaborate to solve diverse parts of a problem, are emerging as the next milestone in artificial intelligence. Here’s how individual systems might work together to solve problems.
- One agent can detect anomalies and analyze telemetry to scope impact, replacing static thresholds with dynamic context-aware evaluation.
- A second agent can iteratively query MELT data to identify root causes, eliminating manual data exploration and hypothesis testing.
- A third agent might execute remediation workflows, adapting actions to real-time system states rather than relying solely on predefined runbooks.
These agents create a cohesive, end-to-end workflow that mimics human expertise while scaling far beyond human capabilities. For observability and incident response, Agentic AI signifies a transition from “task-oriented” workflows to “outcome-oriented” workflows. These agents assess the current system state, plan appropriate actions, interact with relevant tools or data sources, execute tasks, and iteratively refine their strategies based on outcomes. This evolution unlocks a range of possibilities in operationally resilient systems. Agentic AI offers greater ability to navigate expansive telemetry data in alignment with system context and user intent, and can orchestrate remediation workflows that go beyond predefined runbooks, to name just a few of its advantages.
Pre-Agentic vs. Agentic AI in Infrastructure Observability
Let’s do a granular analysis of how Agentic AI systems can evolve IT operations compared to traditional AI.
- Dashboards vs. Autonomous Decision-Making: Pre-agentic workflow systems collect performance metrics (e.g., response times, error rates) and present them to human operators. In contrast, Agentic AI continuously monitors performance metrics and system health indicators (e.g., server load, latency). When it detects anomalies, it flags them and identifies the best course of action based on learned policies.
- Alerting and Recommendations vs. Automated Remediation: Generative AI sends alerts (e.g., via email or chat notifications) to human operators when performance drops below certain thresholds. Agentic AI, however, can automatically scale resources or reroute traffic when it detects potential bottlenecks. It can also restart a service or roll back a deployment in the event of service degradation, with human approvals and guardrails integrated as needed.
- Human-Centric Escalation vs. Predictive Maintenance: In pre-agentic systems, engineers manually investigate issues based on alerts, decide on actions to take, such as rolling back a deployment or reallocating resources, and then execute those actions. Agentic systems, on the other hand, proactively forecast potential failures and schedule maintenance tasks or re-deployments (e.g., expanding blue/green rollout) instead of waiting for thresholds to be violated.
- Limited Autonomy vs. Dynamic Resource Allocation: Pre-agentic AI is largely reactive and depends on user-initiated investigations. Agentic AI adjusts infrastructure in real time based on usage patterns, scaling up during high-traffic periods and scaling down during low-traffic periods.
- Predictable but Slower Response Times vs. Continuous Feedback Loop: Because human intervention is required to deploy fixes or changes, critical issues might be addressed more slowly with pre-agentic AI. In contrast, Agentic AI learns from previous incidents or near misses to refine its decision-making models. Over time, it becomes better at predicting issues and selecting remediation strategies.
How Organizations Can Be Ready for Agentic AI
Organizations hoping to capitalize on advancements in Agentic AI should focus on ensuring that data from various sources is clean, consistent, and well-integrated. They should also invest in developing and hiring talent with expertise in AI, machine learning, and data science, as this will be crucial for maintaining advanced monitoring systems. It’s important that organizations foster a culture of innovation and continuous improvement to encourage teams to experiment with new technologies. Lastly, implementing strong ethical guidelines and security measures will help build trust in these systems.
Interested in the latest in artificial intelligence? Read Sai’s piece outlining the four pillars of AI Observability.
