Sep 8, 2026

Evaluating Agent Performance: Metrics That Matter

Task success rate is a vanity metric. Here are the metrics that actually matter for production agents.

Why Task Success Rate Lies

Imagine an agent that:

Succeeds 90% of the time on simple tasks
Fails catastrophically 10% of the time (deletes database, leaks API keys)
Has no way to recover from failures

Task success rate: 90% → Looks good. Reality: Dangerous.

Production agents need a richer set of metrics.

The Four Essential Metrics

Metric	What It Measures	Why It Matters	Target
Faithfulness	How often the agent sticks to retrieved facts	Prevents hallucinations and made-up answers	≥ 0.9
Hallucination Rate	% of outputs containing unsupported facts	Direct measure of reliability	≤ 0.05
Cost-Per-Task	Average cost (in $) to complete a task	Determines economic viability	< $0.05
Recovery Rate	% of failed tasks that self-recover via retries/failovers	Measures resilience	≥ 0.8

Let’s dive into each.

1. Faithfulness: The Truthfulness Score

Faithfulness measures whether the agent’s output is supported by the sources it consulted.

Definition:
Given a context C (retrieved documents, conversation history) and an output O,
faithfulness = % of claims in O that are entailed by C.

How to Measure:

Extract claims from output O
For each claim, check if it’s supported by context C
Score = (supported claims) / (total claims)

Example:

Context: “The study found magnesium reduced sleep latency by 15 minutes.”
Output: “Magnesium reduces sleep latency by 15 minutes, based on the study.”
Faithfulness: 1.0 (claim is supported)
Output: “Magnesium reduces sleep latency by 30 minutes, based on the study.”
Faithfulness: 0.0 (claim contradicts context)

Tools: Use NLI models or LLM-as-judge to automate.

Target: ≥ 0.9 for production use. Below 0.8 requires immediate investigation.

2. Hallucination Rate: The Reliability Gauge

Hallucination rate measures how often the agent invents facts.

Definition:
% of task outputs that contain at least one hallucinated fact.

How to Measure:

Run agent on a test set with known ground truth
For each output, verify all factual claims against ground truth
Score = (outputs with ≥1 hallucination) / (total outputs)

Types of Hallucinations:

Intrinsic: Contradicts provided context (e.g., context says X, agent says not-X)
Extrinsic: Adds information not in context and not verifiable (e.g., “The study also found Y” when Y wasn’t studied)

Target: ≤ 0.05 (5%) for production. Medical/legal agents: ≤ 0.01.

3. Cost-Per-Task: The Economic Metric

Agents consume tokens and make tool calls. Cost-per-task tells you if it’s affordable.

Definition:
Average total cost (in USD) to complete one task end-to-end.

Calculation:

Cost-Per-Task = 
  (Input Tokens × $/1K_input) + 
  (Output Tokens × $/1K_output) + 
  (Tool Calls × $/call) + 
  (Overhead × $/second)

Example Breakdown for a Research Task:

Input tokens: 2000 × $0.15/1K = $0.30
Output tokens: 500 × $0.60/1K = $0.30
Tool calls: 5 web searches × $0.005 = $0.025
Total: $0.625

Optimization Levers:

Use cheaper models for sub-tasks (Haiku for extraction, Opus for synthesis)
Cache repeated searches and extractions
Compress prompts
Batch tool calls

Target: Depends on use case. For high-volume tasks: < $0.01. For deep research: < $0.50.

4. Recovery Rate: The Resilience Measure

No agent is perfect. Recovery rate measures how well it handles failures.

Definition:
% of failed tasks that the agent successfully recovers from via retries, fallback tools, or self-correction.

How to Measure:

Inject failures (tool errors, timeouts, bad outputs)
Observe agent’s response
Score = (tasks recovered to success) / (total failed tasks)

Recovery Strategies:

Retry with exponential backoff (for transient errors)
Fallback tools (if web_search fails, try cached data)
Self-correction (if output fails validation, regenerate)
Human escalation (after N failures, ask for help)

Example:
Agent tries to extract data from a PDF.

First attempt: PDF corrupted → retry with different parser → success
Second attempt: Parser timeout → fall back to manual extraction → success
Recovery rate: 100% for these failures

Target: ≥ 0.8 for production. Below 0.5 means the agent is brittle.

The Evaluation Dashboard

Track these metrics over time:

# Monitoring configuration
metrics:
  faithfulness:
    window: 1h
    alert: "< 0.85"
    critical: "< 0.7"
  
  hallucination_rate:
    window: 1h
    alert: "> 0.1"
    critical: "> 0.2"
  
  cost_per_task:
    window: 24h
    alert: "> $0.10"
    critical: "> $0.50"
  
  recovery_rate:
    window: 1h
    alert: "< 0.7"
    critical: "< 0.5"

Building a Test Suite

Create a golden dataset for continuous evaluation.

1. Faithfulness Test Set

Context	Query	Expected Output (Supported Claims)
“Study: Mg improves sleep latency by 15 min (N=50)”	“What did the study find about magnesium and sleep?”	“The study found magnesium improved sleep latency by 15 minutes.”
“API returns error 429 when rate limited”	“How should we handle 429 errors?”	“Implement exponential backoff and retry.”

2. Hallucination Test Set

Query	Known Fact	Hallucination to Detect
“What is the capital of France?”	“Paris”	“London”
“What does the study say about dosage?”	“Study tested 200mg”	“Study tested 500mg”

3. Failure Injection Test Set

Tool	Failure Type	Expected Recovery
web_search	Timeout	Retry 2x with backoff
web_extract	404	Fall back to search snippet
LLM call	Rate limit	Queue and retry later

The Takeaway

Stop measuring task success rate. Start measuring what actually matters:

Faithfulness → Does the agent stick to the facts?
Hallucination Rate → How often does it make things up?
Cost-Per-Task → Can you afford to run it at scale?
Recovery Rate → How well does it handle failures?

A production agent isn’t the one that succeeds most often. It’s the one that:

Stays truthful
Admits uncertainty
Manages costs
Bounces back from failures

Measure these. Alert on them. Improve them.
Your users (and your CFO) will thank you.