pygent-test 0.1.2

Pytest-style behavioral regression testing for AI agents.

AgentCheck

AgentCheck is pytest for AI agents. Test behavior, not exact text.

Install from PyPI:

pip install pygent-test

Install from source:

python -m pip install -e .

Optional framework extras:

pip install "pygent-test[langgraph]"
pip install "pygent-test[openai]"

What It Does

AgentCheck helps you verify agent behavior such as:

• which tools were used
• whether tools were used in the expected order
• whether the agent stayed within a step budget
• whether the agent claimed success without tool evidence
• whether behavior regressed against a saved baseline

Current Status

This repo already supports:

• repeated-run behavioral tests with @agent_test(...)
• local baseline and regression comparison
• CLI commands: test, bless, compare, report
• pytest integration
• a plain Python adapter
• an OpenAI Agents SDK adapter
• a LangGraph adapter
• real live OpenAI agent tests in integration_examples/

Quick Start

python -m pip install -e .
python -m agentcheck.cli test examples

Minimal Example

from agentcheck import agent_test, expect
from examples.booking_agent import SimpleBookingAgent


@agent_test(runs=5, agent_factory=SimpleBookingAgent)
def test_booking_agent(agent: SimpleBookingAgent):
    result = agent.run("Book a table for 2 tonight")

    check = expect(result, collect=True)
    check.used_tool("restaurant_search")
    check.used_tool("booking_tool")
    check.steps_less_than(5)
    check.did_not_claim_confirmation_without_tool("booking_tool")
    check.verify()
    return result

Real Agent Testing

AgentCheck has been exercised against:

• real OpenAI Agents SDK agents
• real local LangGraph graphs built with StateGraph

Use the included repo live suite:

python -m agentcheck.cli test integration_examples

or:

python -m pytest integration_examples -q

The included live tests cover:

• a single-tool weather assistant
• a multi-tool research assistant

LangGraph support is tested through the regular unit suite and normalizes the common invoke({"messages": [...]}) flow into AgentResult.

Run the local LangGraph example with:

python -m agentcheck.cli test framework_examples

Documentation

Use these docs depending on what you need:

• TECHNICAL_GUIDE.md Detailed developer guide covering architecture, assertions, adapters, and workflows
• ADAPTER_GUIDE.md How adapters are structured and how to build a new one
• REAL_WORLD_TESTING.md Real OpenAI Agents SDK testing setup and examples
• ROADMAP.md What is done, what is next, and what is planned later

Included Demos

Passing local demo:

python -m agentcheck.cli test examples

Intentional failure demo:

python -m agentcheck.cli test regression_examples --fail-on-regression

Commands

• python -m agentcheck.cli test <path>
• python -m agentcheck.cli bless <path>
• python -m agentcheck.cli compare
• python -m agentcheck.cli report

Smoke Test

If you are working from a source checkout, run a quick end-to-end validation with:

python scripts/smoke_test.py

To include the live OpenAI integration tests:

python scripts/smoke_test.py --with-live

Every agentcheck test run also writes:

• JSON report: .agentcheck/reports/latest.json
• Markdown report: .agentcheck/reports/latest.md

Every agentcheck bless <path> stores a suite-specific baseline under .agentcheck/baselines/.

Baselines are guarded against unrelated suites. If the current suite and saved baseline suite do not match exactly, AgentCheck warns instead of comparing them. For older baseline files without suite metadata, it falls back to matching test names.

Pytest

AgentCheck tests can also run through pytest:

python -m pytest examples -q
python -m pytest tests -q
python -m pytest integration_examples -q

Decorated @agent_test(...) functions are collected as AgentCheck test items, and each item still runs its configured repeated-run behavior.

Assertions

Current built-in assertions:

• used_tool(...)
• used_tool_times(...)
• used_tool_at_least(...)
• used_tool_at_most(...)
• did_not_use_tool(...)
• used_tools_in_order([...])
• steps_less_than(...)
• finished_successfully()
• did_not_error()
• final_output_contains(...)
• final_output_does_not_contain(...)
• did_not_claim_confirmation_without_tool(...)

Use fail-fast assertions:

expect(result).used_tool("restaurant_search")

Use collected assertions when you want one run to report multiple failures:

check = expect(result, collect=True)
check.used_tool("restaurant_search")
check.used_tool("booking_tool")
check.did_not_claim_confirmation_without_tool("booking_tool")
check.verify()

Roadmap

This is the first step.

Near-term priorities:

• cleaner regression summaries
• better onboarding for testing a real agent in under 5 minutes
• more adapters based on actual user demand

Longer-term directions:

• stronger regression analysis
• better flakiness reporting
• richer CI workflows
• optional hosted features only if the core library proves valuable

For a more detailed breakdown, see ROADMAP.md.