Agent Evaluation Methods

Agent evaluation requires different approaches than traditional software. Agents are non-deterministic, may take different valid paths, and lack single correct answers.

Key Finding: 95% Performance Drivers

Research on BrowseComp found three factors explain 95% of variance:

Factor	Variance	Implication
Token usage	80%	More tokens = better performance
Tool calls	~10%	More exploration helps
Model choice	~5%	Better models multiply efficiency

Implications: Model upgrades beat token increases. Multi-agent architectures validate.

Multi-Dimensional Rubric

Dimension	Excellent	Good	Acceptable	Failed
Factual accuracy	All correct	Minor errors	Some errors	Wrong
Completeness	All aspects	Most aspects	Key aspects	Missing
Citation accuracy	All match	Most match	Some match	Wrong
Tool efficiency	Optimal	Good	Adequate	Wasteful

LLM-as-Judge

evaluation_prompt = """
Task: {task_description}
Agent Output: {agent_output}
Ground Truth: {ground_truth}

Evaluate on:
1. Factual accuracy (0-1)
2. Completeness (0-1)
3. Citation accuracy (0-1)
4. Tool efficiency (0-1)

Provide scores and reasoning.
"""

Test Set Design

test_set = [
    {"name": "simple", "complexity": "simple",
     "input": "What is capital of France?"},
    {"name": "medium", "complexity": "medium",
     "input": "Compare Apple and Microsoft revenue"},
    {"name": "complex", "complexity": "complex",
     "input": "Analyze Q1-Q4 sales trends"},
    {"name": "very_complex", "complexity": "very_complex",
     "input": "Research AI tech, evaluate impact, recommend strategy"}
]

Evaluation Pipeline

def evaluate_agent(agent, test_set):
    results = []
    for test in test_set:
        output = agent.run(test["input"])
        scores = llm_judge(output, test)
        results.append({
            "test": test["name"],
            "scores": scores,
            "passed": scores["overall"] >= 0.7
        })
    return results

Complexity Stratification

Level	Characteristics
Simple	Single tool call
Medium	Multiple tool calls
Complex	Many calls, ambiguity
Very Complex	Extended interaction, deep reasoning

Context Engineering Evaluation

Test context strategies systematically:

Run agents with different strategies on same tests
Compare quality scores, token usage, efficiency
Identify degradation cliffs at different context sizes

Continuous Evaluation

Run evaluations on all agent changes
Track metrics over time
Set alerts for quality drops
Sample production interactions

Avoiding Pitfalls

Pitfall	Solution
Path overfitting	Evaluate outcomes, not steps
Ignoring edge cases	Include diverse scenarios
Single metric	Multi-dimensional rubrics
Ignoring context	Test realistic context sizes
No human review	Supplement automated eval

Best Practices

Use multi-dimensional rubrics
Evaluate outcomes, not specific paths
Cover complexity levels
Test with realistic context sizes
Run evaluations continuously
Supplement LLM with human review
Track metrics for trends
Set clear pass/fail thresholds

agent-evaluation