AI Agents Explained for SRE & DevOps Engineers

An agent is not just an LLM

An Agent is an application pattern where a model can choose actions, call tools, observe results and decide what to do next. The important operational difference is that one user request can create a variable number of downstream actions.

Goal
 ↓
LLM decides next action
 ↓
Tool call
 ↓
Observation
 ↓
Decide again
 ↓
Stop condition / answer

Workflow vs agent

A Workflow follows steps defined ahead of time. An agent has more freedom to choose its next step. Read Workflow vs Agent whenever someone describes every automation as “agentic.” The distinction matters because dynamic behavior changes reliability, cost and security boundaries.

What can fail

• Tool timeout or bad response.
• Agent Loop that continues too long.
• Unexpected tool sequence or excessive agency.
• Prompt injection causing unsafe tool choice.
• Token and API cost multiplying across iterations.
• Partial state after some actions succeed and others fail.

How an SRE should think about an agent

Treat the agent loop as a distributed workflow whose control flow is partly probabilistic. Trace every model decision and tool call, cap iterations and budgets, make important actions idempotent, and define clear timeouts and stop conditions.

Where MCP fits

Model Context Protocol (MCP) can standardize how applications connect models to tools and resources. It does not make tools safe automatically. Authentication, least privilege, validation and auditability still belong to the surrounding production system.