LLM evaluation tests that actually mean something. A pytest plugin for testing LLM agents with threshold-based pass/fail scoring, multi-turn transcripts, and LLM-as-judge rubrics — without breaking your CI bill.
- 🎯 Threshold-based pass/fail — run each test N times, pass when ≥ threshold% succeed
- 📝 YAML or Python transcripts — pick the authoring style your team prefers
- 🔍 YAML auto-discovery — drop
*.yamlfiles in any configured directory and they become pytest tests automatically - 🎙 Voice agents (LiveKit) — drive a real
AgentSessionwith a WAV per turn; same evaluator surface as text agents - 🚦 Group-level quality gates —
[tool.agent_eval.groups]gates CI on aggregate pass rates ("90% of booking evals") withmust_passpins, instead of failing on every flaky eval - 🛡 CI-safe by default — eval tests skip unless
--agent-eval-liveorEVAL_LIVE=1 - ⚡ Parallel-ready —
pytest -n auto(viapytest-xdist) just works - 📄 Markdown reports — full per-run trace with
--agent-eval-report=eval.md
# pip
pip install pytest-agent-eval
# uv
uv add pytest-agent-evalpytest-agent-eval ships first-class adapters for the major Python agent frameworks. Each is an optional extra so you only install what you use.
| Framework | Extra | Adapter |
|---|---|---|
| pydantic-ai | (default) | pytest_agent_eval.adapters.pydantic_ai.PydanticAIAdapter |
| LangChain / LangGraph | langchain |
pytest_agent_eval.adapters.langchain.LangChainAdapter |
| OpenAI SDK | openai |
pytest_agent_eval.adapters.openai.OpenAIAdapter |
| smolagents | smolagents |
pytest_agent_eval.adapters.smolagents.SmolagentsAdapter |
| LiveKit (voice) | livekit |
pytest_agent_eval.adapters.livekit.LiveKitAdapter |
pip install "pytest-agent-eval[langchain]"
pip install "pytest-agent-eval[openai]"
pip install "pytest-agent-eval[smolagents]"
pip install "pytest-agent-eval[livekit]"
# or with uv:
uv add "pytest-agent-eval[langchain]"
uv add "pytest-agent-eval[openai]"
uv add "pytest-agent-eval[smolagents]"
uv add "pytest-agent-eval[livekit]"Bringing your own framework? Any async def agent(history) -> AgentReply callable works directly — no base class needed. A plain (reply, tool_calls) tuple is still accepted.
pytest-agent-eval separates the kinds of checks you might want into composable evaluators:
- Deterministic checks —
ContainsEvaluator(any_of=["confirmed", "booked"], matches_all=[r"BK-\d+"])for substring and regex assertions over the agent reply. - Tool-call assertions —
ToolCallEvaluator(must_include=["create_booking"], ordered=True)to verify that the agent called the right tools, in the right order. - Tool-argument assertions —
ToolCallArgsEvaluator(tool="create_booking", args={"time": "10am"})for deterministic subset/exact checks on call arguments, plusToolCallArgsJudgeEvaluatorfor rubric-based judging of arguments. - LLM-as-judge —
JudgeEvaluator(rubric="Reply must be friendly, include a date, and confirm the booking.")for open-ended quality checks the agent under test should meet.
Mix and match per turn — every evaluator participates in the threshold score.
You can author evals two ways. YAML is the recommended starting point — it lets non-Python contributors (PMs, QA, domain experts) write tests, keeps eval data readable in code review, and turns each .yaml file into a pytest test automatically.
# pyproject.toml
[tool.agent_eval]
model = "openai:gpt-4o" # used by the LLM-as-judge
threshold = 0.8 # default pass fraction
runs = 3 # default reps per transcript
yaml_dirs = ["tests/evals"]# tests/conftest.py
import pytest
from pytest_agent_eval.adapters.pydantic_ai import PydanticAIAdapter
from my_app import build_agent
@pytest.fixture
def llm_eval_agent():
return PydanticAIAdapter(build_agent())# tests/evals/booking_single_turn.yaml
# yaml-language-server: $schema=https://datarootsio.github.io/pytest-agent-eval/schema/transcript.json
id: booking_single_turn
threshold: 0.8
runs: 3
turns:
- user: "Book me a slot tomorrow at 10am."
expect:
reply_contains_any: ["confirmed", "booked"]
tool_calls_include: ["create_booking"]
judge:
rubric: "Reply must confirm the booking and include a reference number."pytest --agent-eval-live============================ test session starts =============================
plugins: agent-eval-0.1.0, asyncio-1.0.0
collected 2 items
tests/evals/booking_single_turn.yaml::booking_single_turn PASSED [ 50%]
tests/evals/booking_multi_turn.yaml::booking_multi_turn PASSED [100%]
============================== 2 passed in 14.03s ============================
By default eval tests are skipped outside of explicit live runs (so a missed pytest . doesn't burn API credits). Flip --agent-eval-live on, set EVAL_LIVE=1, or live = true in [tool.agent_eval] for local-only auto-on. See Configuration for the full precedence rules.
Real agents fail on context, not single-shot replies. A model that nails turn 1 might forget the user's name by turn 3, or call the wrong tool once the user changes their mind. Multi-turn YAML transcripts test the whole conversation arc, with each turn asserting against the agent's state at that point:
# tests/evals/booking_multi_turn.yaml
id: booking_multi_turn
threshold: 0.66 # tolerate 1/3 flaky runs
runs: 3
tags: [gate:booking, smoke]
turns:
# Turn 1 — initial booking
- user: "Book me a slot tomorrow at 10am."
expect:
reply_contains_any: ["confirmed", "booked"]
tool_calls_include: ["create_booking"]
# Turn 2 — agent must remember the booking from turn 1
- user: "Actually, can you move it to 11am instead?"
expect:
tool_calls_include: ["update_booking"]
tool_calls_exclude: ["create_booking"] # must update, not double-book
judge:
rubric: "Confirms the new time AND references the original 10am booking."
# Turn 3 — context propagates further
- user: "Email me the confirmation."
expect:
tool_calls_include: ["send_email"]
reply_contains_any: ["sent", "email"]Each turn's full conversation history is built up as the test runs — your agent receives all prior (user, assistant) pairs as context, the same way it would in production. Failures point at the exact turn that broke, not just "the test failed."
A suite of LLM evals fails somewhere almost every run. Group thresholds gate CI on aggregate pass rates while pinning the tests that must never break:
[tool.agent_eval.groups.booking]
threshold = 0.9 # 90% of matched tests must pass
tags = ["gate:booking"] # match transcripts by tag
must_pass = ["booking_confirmation"] # ...but this one must always pass============================== group summary ===============================
booking: 9/10 passed (90%) >= 90% required -- PASSED
failures: booking_edge_case
must_pass: booking_confirmation ok
exit code overridden to 0: all group thresholds met
When every gate is green and every failure belongs to a gated group, the exit code is overridden to 0 — a failing plain unit test or ungrouped transcript still keeps CI red. See the group thresholds docs for membership and must_pass semantics.
The [livekit] extra adds a LiveKitAdapter that drives a real LiveKit AgentSession from a WAV per turn. Every turn declares an audio: path — the adapter streams it at real-time pace, captures function_tools_executed and conversation_item_added events, and returns (reply, tool_calls) to the same evaluators you already use for text agents:
# tests/evals/booking_voice.yaml
id: booking_voice
turns:
- user: "Book me a slot tomorrow at 10am."
audio: booking_t1.wav # resolved relative to this YAML's directory
expect:
tool_calls_include: [create_booking]
reply_contains_any: [confirmed, booked]# tests/conftest.py
from livekit.agents.voice import Agent, AgentSession
from livekit.plugins import openai
from pytest_agent_eval.adapters.livekit import LiveKitAdapter
def make_session():
session = AgentSession(llm=openai.realtime.RealtimeModel())
agent = Agent(instructions="You are a booking assistant.", tools=[...])
return session, agent
@pytest.fixture
def llm_eval_agent():
return LiveKitAdapter(make_session)A bundled CLI generates the WAVs from each turn's user: text via OpenAI Realtime — hash-cached so unchanged transcripts skip re-synthesis, and idempotent enough for CI prebuilds:
python -m pytest_agent_eval.synthesize_audio # walks [tool.agent_eval].yaml_dirs
python -m pytest_agent_eval.synthesize_audio tests/evals/ # explicit dir
python -m pytest_agent_eval.synthesize_audio --force # ignore cacheThe CLI auto-writes a .gitignore next to every WAV (*.wav, *.wav.hash) so generated audio stays local — commit YAML transcripts only. Real recordings work too: drop a hand-recorded WAV at the same path and the adapter doesn't care how it got there.
See Voice testing for the full reference (sample rate, frame size, grace period, custom session factories).
Transcripts are plain YAML with a published JSON Schema, validation errors suggest the intended field ("Did you mean 'tool_calls_include'?"), and the docs ship as llms.txt / llms-full.txt — so Claude Code, Cursor, and friends write correct evals on the first try. Grab the ready-made AGENTS.md snippet from the coding agents guide, and point agents at examples/ for known-good starting points.
Add --agent-eval-report=eval.md to get a human-readable trail of every run, every turn, and every evaluator's reasoning. Useful for CI artifacts and PR diffs:
pytest --agent-eval-live --agent-eval-report=eval.md# LLM Eval Report — 2026-04-30
## Summary
| Transcript | Runs | Passed | Score | Threshold | Status |
|-----------------------|------|--------|-------|-----------|---------|
| booking_single_turn | 3 | 3 | 1.00 | 0.80 | ✅ PASS |
| booking_multi_turn | 3 | 2 | 0.67 | 0.66 | ✅ PASS |
## Details
### booking_multi_turn
**Run 1** ✅
- Turn 1: PASS
- Turn 2: PASS
- Judge: Reply confirmed move to 11am and acknowledged original 10am slot.
- Turn 3: PASS
**Run 2** ❌
- Turn 1: PASS
- Turn 2: FAIL
- Tool calls expected to include 'update_booking', got ['create_booking']
- Turn 3: PASS
**Run 3** ✅
- ...The two-out-of-three pass still clears the 0.66 threshold, so the suite passes — that's the point of running each transcript multiple times instead of treating LLM tests as binary.
YAML covers most cases; drop into Python when you need parametrization, programmatic test generation, or per-test fixtures:
import pytest
from pytest_agent_eval import Turn, Expect, ContainsEvaluator, ToolCallEvaluator, JudgeEvaluator
@pytest.mark.agent_eval(threshold=0.8, runs=3)
async def test_booking(agent_eval):
result = await agent_eval.run(
agent=my_agent,
turns=[
Turn(
user="Book me a slot tomorrow at 10am",
expect=Expect(evaluators=[
ContainsEvaluator(any_of=["confirmed", "booked"]),
ToolCallEvaluator(must_include=["create_booking"]),
JudgeEvaluator(rubric="Reply must include a reference number."),
]),
)
],
)
result.assert_threshold()See the full documentation (or /main/ for the in-development version) for the complete YAML reference, configuration options, parallel execution, and CI patterns. Contributions are welcome — see CONTRIBUTING.md.
MIT — see LICENSE.