There's a saying in AI circles: monitoring tells you what's happening — evaluation tells you how good it is. You can have an agent that responds instantly, never crashes, and answers every question with absolute confidence. But confidence without correctness is just a well-dressed mistake.
That's exactly why Oracle AI Agent Studio gives us two complementary capabilities: Monitoring and Evaluation. In this post, we're focusing on Evaluation — what it is, how to set it up, and why you should care before your agent goes anywhere near production.
Curious about Monitoring? Check out this post.
Why Evaluation Matters
Evaluation ensures your agent can meet defined standards and business outcomes across three dimensions: accuracy, latency, and token usage.
Without it, you're essentially deploying on vibes.
With a proper evaluation framework, you can validate that your agent answers correctly, responds within acceptable time, and doesn't burn through your token budget faster than your cloud spend budget. It's the closest thing to a test suite for your AI agent.
The Metrics: What Gets Measured
Oracle AI Agent Studio provides a rich set of metrics, and it's worth understanding which ones are available for Evaluation versus Monitoring, because they serve different purposes.
| Metric | Evaluation | Monitoring |
|---|---|---|
| Error Rate | ✅ | ✅ |
| Error Count | ✅ | ✅ |
| Session Count | ✅ | ✅ |
| P99 Latency | ✅ | ✅ |
| P50 Latency | ✅ | ✅ |
| Total Tokens | ✅ | ✅ |
| Input Token Count | ❌ | ✅ |
| Output Token Count | ✅ | ✅ |
| Median Correctness | ✅ | ❌ |
| Groundedness | ✅ | ❌ |
| Answer Relevance | ✅ | ❌ |
| Context Relevance | ✅ | ❌ |
The quality metrics — Correctness, Groundedness, Answer Relevance, and Context Relevance — are exclusive to Evaluation. These are the metrics that tell you whether your agent is genuinely useful, not just technically operational.
A quick breakdown of what these mean in practice:
Median Correctness measures how closely your agent's answer matches the expected reference response. Scores range from 0 to 1. Think of it as your agent's grade on the test.
Groundedness measures whether the generated answer is actually grounded in the retrieved source content. A grounded response stays faithful to what the knowledge base says — it doesn't hallucinate or embellish. (Hallucination: the nemesis of every enterprise AI implementation.)
Answer Relevance measures how directly and precisely the agent's response addresses the user's question. Getting the right answer to the wrong question doesn't count.
Context Relevance measures the quality of the retrieved information itself — whether the context pulled in by the agent was actually appropriate and reliable enough to produce a good answer.
Setting Up an Evaluation: Step by Step
Step 1 — Define Your Evaluation Set
Before you can run anything, you need to define an evaluation set. Think of this as your test plan. It includes:
- Test questions — the inputs your agent needs to handle
- Expected responses — the gold-standard answers you're measuring against
- Success criteria — the thresholds each metric needs to meet
An evaluation set without expected responses is just a demo. The expected responses are what turn a run into a meaningful quality gate.
Step 2 — Choose Your Run Mode
When setting up your evaluation run, you'll choose between two modes:
Sequential runs questions in the exact order you define them. Use this when one question depends on the context from the previous one — for example, a multi-turn conversation flow.
Random runs questions in a randomised order. This is useful when testing independent questions where order doesn't matter, and it helps reduce positional bias in your evaluation results.
Step 3 — Define Your Questions
In the Questions tab, add the questions users are expected to ask your agent, paired with the exact responses you want the agent to return.
Here's an example from an HR benefits agent:
Q: Who is eligible for the benefits program?
A: Eligibility Criteria:
- Full-time employees working 30+ hours per week are eligible for full benefits.
- Part-time employees may qualify for limited benefits.
- Benefits eligibility begins on the first day of the month following hire date.
Dependents (spouse and children under age 26) may be enrolled in applicable plans.
Keep your expected responses as close to production-quality as possible. The correctness metric is only as good as the reference answer you define.
Step 4 — Configure Your Metrics
In the Metrics tab, you select which metrics to include in this evaluation run. This is where you tailor the evaluation to your agent's specific use case and business requirements.
For example:
- If your agent doesn't invoke any APIs, you can exclude API error metrics — no point cluttering your results with noise.
- If accuracy is your top priority (say, a policy or compliance agent), set your correctness threshold high —
0.8is a reasonable baseline for enterprise use. - If token cost is a concern, configure output token thresholds to flag responses that are running unnecessarily long.
Metrics without thresholds are just numbers. Thresholds are what turn numbers into pass/fail signals.
Step 5 — Initiate the Evaluation Run
Click Initiate Evaluation Run. Oracle AI Agent Studio will execute the evaluation and return results for each question, including:
- Actual response vs. expected response — side by side
- Latency per question
- Token usage (input and output)
- Quality scores for the metrics you selected
Reading Your Results
After the run completes, reviewing the results is where the real value surfaces. Here's an example of what you might find:
Latency: One question took over 20 seconds — exceeding the defined threshold. That's a red flag worth investigating. It could point to an overly complex retrieval step, a large system prompt, or a knowledge base that needs optimisation. The remaining questions came in well within threshold.
Token Usage: Both input and output token counts were within acceptable limits. Good news for the budget.
Correctness: With a threshold of 0.8, any question scoring below that benchmark gets flagged for review. Patterns in low-scoring questions often reveal gaps in your knowledge base or ambiguities in your system prompt.
This combination of latency, cost, and quality signals gives you a complete picture — not just "did it answer?" but "did it answer well, quickly, and efficiently?"
Final Thoughts
An AI agent that passes evaluation isn't just technically sound — it's one you can actually stand behind when a business user asks, "How do we know this is right?"
Defining quality thresholds, building meaningful evaluation sets, and reviewing results against expected outcomes is what separates a production-ready agent from a prototype running in a demo environment. Oracle AI Agent Studio gives you the tooling to do this properly. Use it.
Top comments (0)