AgentProof

pytest-based behavioral testing for AI agents.

pip install agentproof

One line. No config. Works with any agent framework.

The Problem

You can observe your agents. You can trace them. You can log every token.

But can you test them?

89% of teams have agent observability. Only 52% have agent evaluation. Zero have behavioral testing in CI.

Your agent calls the wrong tool? Ships to production. Costs $40 on a $0.50 task? Ships to production. Hallucinates the answer? Ships. To. Production.

The Solution

AgentProof brings pytest-style behavioral testing to AI agents. Test what your agent does, not just what it outputs.

# test_booking_agent.py
import agentproof
from agentproof import assert_tool_called, assert_tool_order, assert_max_cost

def test_booking_agent_searches_before_booking(agent_run):
    run = make_booking_run()  # your agent execution

    # Did it use the right tools?
    assert_tool_called(run, "search_flights")
    assert_tool_called(run, "book_flight")

    # Did it use them in the right order?
    assert_tool_order(run, ["search_flights", "compare_prices", "book_flight"])

    # Did it stay within budget?
    assert_max_cost(run, max_usd=0.50)

$ pytest
========================= test session starts =========================
test_booking_agent.py::test_booking_agent_searches_before_booking PASSED
test_booking_agent.py::test_cost_stays_under_budget PASSED
test_booking_agent.py::test_no_hallucination PASSED
========================= 3 passed in 0.04s ============================

Quickstart

1. Install

pip install agentproof

2. Record an agent run

import agentproof

@agentproof.record
def my_agent(prompt: str) -> str:
    # Your agent code here
    agentproof.add_tool_call("search", arguments={"query": prompt})
    agentproof.add_llm_call("gpt-4o", prompt_tokens=500, completion_tokens=200)
    return "The answer is 42"

result = my_agent("What is the meaning of life?")
run = my_agent.last_run

3. Write tests

from agentproof import (
    assert_tool_called,
    assert_tool_order,
    assert_max_cost,
    assert_max_steps,
    assert_no_hallucination,
)

def test_my_agent():
    run = my_agent.last_run

    assert_tool_called(run, "search")
    assert_max_cost(run, 0.10)
    assert_max_steps(run, 5)
    assert_no_hallucination(run, ground_truth="42")

Core Assertions

Framework Adapters

Works with any agent framework. Install the adapter you need:

pip install agentproof[langchain]
pip install agentproof[crewai]
pip install agentproof[openai]
pip install agentproof[otel]      # Any OTEL-instrumented framework

# LangChain
from agentproof.adapters.langchain import from_langchain_run
run = from_langchain_run(agent_executor.invoke({"input": "..."}))

# CrewAI
from agentproof.adapters.crewai import from_crewai_result
run = from_crewai_result(crew.kickoff())

# OpenAI Agents SDK
from agentproof.adapters.openai_agents import from_openai_response
run = from_openai_response(Runner.run(agent, "..."))

Replay Testing

Record once, test forever. Capture a real agent run and replay it in CI without making API calls:

from agentproof import save_trace, load_trace

# Record
save_trace(run, "tests/fixtures/booking_success.jsonl")

# Replay in tests
def test_booking_replay():
    run = load_trace("tests/fixtures/booking_success.jsonl")
    assert_tool_called(run, "book_flight")
    assert_max_cost(run, 0.50)

Snapshot Testing

Like Jest snapshots but for agent behavior. Detect behavioral regressions automatically:

from agentproof import assert_snapshot

def test_agent_behavior_stable(snapshot_dir):
    run = my_agent("book a flight to NYC")
    assert_snapshot(run, snapshot_dir / "booking_agent.json")
    # First run: creates snapshot
    # Subsequent runs: compares against saved snapshot

Update snapshots: pytest --agentproof-update-snapshots

Bundled Cost Database

200+ LLM models with up-to-date pricing. No API calls needed.

from agentproof import get_model_pricing, calculate_run_cost

pricing = get_model_pricing("gpt-4o")
# {"provider": "openai", "prompt": 2.50, "completion": 10.00}

cost = calculate_run_cost(run)
# CostBreakdown(total_cost_usd=0.0325, ...)

Models: GPT-4o, Claude 4.5, Gemini 2.5, Llama 4, DeepSeek V3, Mistral Large, and 190+ more.

CI/CD Integration

GitHub Actions

- uses: praxiumlabs/agentproof@v1
  with:
    test-path: tests/

Or manually:

- run: |
    pip install agentproof
    pytest tests/ -v

Comparison

Honest Limitations

• Hallucination detection uses TF-IDF — it catches obvious fabrication, not subtle inaccuracies. For LLM-as-judge evaluation, use DeepEval or Braintrust.
• Token counts must be provided — we don't intercept API calls (by design). Use framework adapters or log tokens manually.
• No real-time monitoring — AgentProof is a testing tool, not an observability platform. Use it alongside your existing tracing.
• Cost DB needs community updates — model pricing changes. Submit a PR to update agentproof/data/models.yaml.

Contributing

git clone https://github.com/praxiumlabs/agentproof
cd agentproof
pip install -e ".[dev]"
pytest

Adding a model to the cost DB:

Edit agentproof/data/models.yaml:

new-model-name:
  provider: provider-name
  prompt: 1.50
  completion: 5.00

License

MIT License. See LICENSE.

Built by Sarak Dahal at Praxium Labs. Star the repo if AgentProof saves your agents.