Introduction
The Challenge
When an agent gets it wrong #
Imagine a patient posts this the day after your assistant goes live:
“I tried the new AI assistant to ask about my medication, and it told me to take double the dose listed on my prescription. That’s not a small mistake. That’s dangerous. If I hadn’t double-checked with my doctor, I could have actually followed that advice.”
The infrastructure was healthy. The API returned 200 OK. Latency was fine. Every
traditional signal said the system was working, and yet the agent produced an answer that
could harm someone. That’s the agentic AI trust gap, and it’s invisible to
infrastructure and APM monitoring alone.
Agentic systems introduce a new class of risk:
- Unpredictable reasoning: the agent chooses a path you didn’t anticipate.
- Hallucinations: confident answers that aren’t grounded in your data.
- Tool-selection errors: the agent calls the wrong tool, or the right tool with the wrong arguments.
- Cost spikes: runaway loops and oversized prompts quietly burning tokens.
- Sensitive-data exposure: the agent reveals or acts on data it shouldn’t.
What Splunk Agent Observability gives you #
Splunk Agent Observability, powered by Galileo, is built to close that gap across the entire agent lifecycle, from development through production.
Three things you get
- Accurate evaluations: evaluate agent, output, and RAG quality with high-accuracy evals plus out-of-the-box and custom evaluators, so “is this answer good?” becomes a number you can track.
- Instant visibility: see the root cause of errors across complex, multi-step agent workflows, instead of guessing from logs.
- Runtime guardrails: block or steer hallucinations, prompt injections, and safety violations before they ever reach an end user.
Over the rest of this workshop you’ll apply all three to the Careful Health Provider assistant, turning that frightening social post into a problem you can detect, measure, prevent, and govern.
The assistant told a patient to take double their dose, but every infrastructure and APM signal was green. Why didn’t traditional monitoring catch it?