AI Agent Testing Strategies: Unit, Integration, and End-to-End Approaches

The Testing Problem Is Different for Agents

Traditional software testing relies on deterministic behavior: given input X, expect output Y. AI agents introduce non-determinism at their core — the same input can produce different outputs, different tool call sequences, and different reasoning paths. This does not mean agents are untestable. It means we need a testing framework designed for probabilistic systems.

A practical agent testing strategy operates at three levels, each catching different categories of defects.

Level 1: Unit Tests (Deterministic)

Unit tests validate the deterministic components of your agent system — everything except the LLM calls themselves.

flowchart LR
    PR(["PR opened"])
    UNIT["Unit tests"]
    EVAL["Eval harness<br/>PromptFoo or Braintrust"]
    GOLD[("Golden set<br/>200 tagged cases")]
    JUDGE["LLM as judge<br/>plus regex graders"]
    SCORE["Aggregate score<br/>and per slice"]
    GATE{"Score regress<br/>more than 2 percent?"}
    BLOCK(["Block merge"])
    MERGE(["Merge to main"])
    PR --> UNIT --> EVAL --> GOLD --> JUDGE --> SCORE --> GATE
    GATE -->|Yes| BLOCK
    GATE -->|No| MERGE
    style EVAL fill:#4f46e5,stroke:#4338ca,color:#fff
    style GATE fill:#f59e0b,stroke:#d97706,color:#1f2937
    style BLOCK fill:#dc2626,stroke:#b91c1c,color:#fff
    style MERGE fill:#059669,stroke:#047857,color:#fff

What to Unit Test

• Tool functions: Each tool the agent can call should have standard unit tests with known inputs and expected outputs
• State management: State transitions, reducers, and serialization logic
• Input validation: Prompt template rendering, parameter parsing, and guardrail logic
• Output parsing: Extracting structured data from LLM responses

# Test a tool function deterministically
def test_calculate_shipping_cost():
    result = calculate_shipping(weight_kg=2.5, destination="US", method="express")
    assert result["cost"] == 24.99
    assert result["estimated_days"] == 3

# Test output parsing
def test_parse_agent_action():
    raw_response = "I'll look up the order. ACTION: get_order(order_id='ORD-123')"
    action = parse_action(raw_response)
    assert action.tool == "get_order"
    assert action.params == {"order_id": "ORD-123"}

Mock LLM Responses

For unit testing agent control flow, replace the LLM with deterministic mock responses:

class MockLLM:
    def __init__(self, responses: list[str]):
        self.responses = iter(responses)

    async def generate(self, prompt: str) -> str:
        return next(self.responses)

# Test the agent's decision logic with predictable LLM outputs
async def test_agent_routes_to_billing():
    mock = MockLLM(["The customer is asking about billing."])
    agent = SupportAgent(llm=mock)
    result = await agent.classify("Why was I charged twice?")
    assert result.category == "billing"

Level 2: Integration Tests (Semi-Deterministic)

Integration tests verify that agent components work together correctly, including interactions with external tools and services.

What to Integration Test

• Tool orchestration: Does the agent call tools in a valid sequence?
• Error handling: Does the agent recover gracefully from tool failures?
• Guardrail enforcement: Do safety checks prevent unauthorized actions?
• State persistence: Does checkpointing and recovery work correctly?

Strategies for Reducing Non-Determinism

• Fixed seeds and low temperature: Set temperature to 0 and use fixed random seeds to increase reproducibility
• Assertion on patterns, not exact text: Check that the agent called the right tools with the right parameters, not that it phrased its reasoning identically
• Bounded retries: Allow tests to retry up to 3 times, passing if any attempt succeeds (for truly non-deterministic outputs)

Level 3: End-to-End Evaluation (Probabilistic)

E2E tests run the full agent pipeline with real LLM calls against a suite of test scenarios. These tests are evaluated probabilistically rather than with exact assertions.

LLM-as-Judge Pattern

Use a separate LLM to evaluate whether the agent's response meets quality criteria:

async def evaluate_response(scenario, agent_response):
    eval_prompt = f"""
    Scenario: {scenario.description}
    Expected behavior: {scenario.expected_behavior}
    Agent response: {agent_response}

    Rate the agent's response on these criteria (1-5):
    1. Correctness: Did it solve the problem?
    2. Completeness: Did it address all aspects?
    3. Safety: Did it stay within authorized boundaries?
    4. Tone: Was the communication appropriate?

    Return JSON: {{"correctness": N, "completeness": N, "safety": N, "tone": N}}
    """
    return await eval_llm.generate(eval_prompt)

Test Scenario Design

Build a diverse evaluation dataset covering:

• Happy paths: Common requests the agent should handle well
• Edge cases: Unusual inputs, ambiguous requests, multi-step problems
• Adversarial inputs: Prompt injections, out-of-scope requests, attempts to bypass guardrails
• Regression cases: Specific failures from production that have been fixed

Setting Pass Thresholds

• Track aggregate scores across the full test suite, not individual scenarios
• Set minimum thresholds (e.g., average correctness above 4.0 out of 5.0)
• Monitor score trends over time to catch gradual degradation

CI/CD Integration

• Unit tests: Run on every commit. Fast, deterministic, no API costs.
• Integration tests: Run on pull requests. Moderate speed, minimal API costs with mock LLMs.
• E2E evaluation: Run nightly or on release candidates. Slow, involves real API costs.

The goal is not to make agent behavior perfectly deterministic — it is to build confidence that the agent handles the scenarios your users encounter, with quality that meets your standards.

Sources: DeepEval Testing Framework | LangSmith Evaluation | Braintrust AI Evaluation