evaldeck 0.1.4

The evaluation framework for AI agents. Pytest for agents.

Evaldeck

The evaluation framework for AI agents. Pytest for agents.

---

Evaldeck helps you answer one question: "Is my agent actually working?"

Unlike LLM evaluation tools that focus on input→output scoring, Evaldeck evaluates the entire agent execution—how it reasons, which tools it selects, and whether it achieves the goal.

Why Evaldeck?

• Agent-native: Evaluates multi-step traces, not just final outputs
• Framework-agnostic: Works with LangChain, CrewAI, AutoGen, or custom agents
• Developer-friendly: CLI-first, CI/CD ready, 5-minute setup
• Comprehensive metrics: Tool correctness, step efficiency, plan adherence, and more
• Flexible grading: Code-based, model-based (BYOK), or combine both

Installation

pip install evaldeck

With framework integrations:

pip install evaldeck[langchain]  # LangChain/LangGraph support
pip install evaldeck[openai]     # OpenAI model graders
pip install evaldeck[all]        # Everything

Quick Start

1. Initialize your project

evaldeck init

This creates:

evaldeck.yaml          # Configuration
tests/evals/         # Test directory
  example.yaml       # Example test case

2. Define test cases

# tests/evals/booking.yaml
name: book_flight_basic
input: "Book me a flight from NYC to LA on March 15th"

expected:
  tools_called:
    - search_flights
    - book_flight
  output_contains:
    - "confirmation"
    - "March 15"
  max_steps: 5

3. Run evaluations

evaldeck run

Output:

Running 3 tests...

  ✓ book_flight_basic (1.2s)
  ✓ book_flight_roundtrip (2.1s)
  ✗ book_flight_with_preferences (1.8s)
    └─ FAIL at step 3: Wrong tool called
       Expected: search_flights_with_filters
       Got: search_flights

Results: 2/3 passed (66.7%)

Configuration

# evaldeck.yaml
version: 1

agent:
  module: my_agent
  function: run

test_dir: tests/evals

defaults:
  timeout: 30
  retries: 2

graders:
  llm:
    model: gpt-4o-mini
    # Uses OPENAI_API_KEY from environment

thresholds:
  min_pass_rate: 0.9

Test Case Format

Basic test case

name: test_name
input: "User message to the agent"

expected:
  # What tools should be called?
  tools_called:
    - tool_name_1
    - tool_name_2

  # What tools should NOT be called?
  tools_not_called:
    - dangerous_tool

  # What should the output contain?
  output_contains:
    - "expected phrase"

  # What should the output NOT contain?
  output_not_contains:
    - "error"

  # Maximum steps allowed
  max_steps: 10

  # Must complete successfully?
  task_completed: true

Using model-based grading

name: helpful_response
input: "Explain quantum computing"

graders:
  - type: llm
    prompt: |
      Rate this response for helpfulness and accuracy.
      Response: {{ output }}

      Score from 1-5, where 5 is excellent.
    threshold: 4

Framework Integration

LangChain

Copy the reference tracer from examples/langchain_tracer.py to your project:

from langchain_tracer import EvaldeckTracer
from langchain.agents import AgentExecutor

tracer = EvaldeckTracer()
agent = AgentExecutor(...)

result = agent.invoke(
    {"input": "Book a flight"},
    config={"callbacks": [tracer]}
)

# Get trace for evaluation
trace = tracer.get_trace()

Manual trace construction

from evaldeck import Trace, Step, Evaluator

trace = Trace(
    input="Book a flight from NYC to LA",
    steps=[
        Step(
            type="tool_call",
            tool_name="search_flights",
            tool_args={"from": "NYC", "to": "LA"},
            tool_result=[{"flight": "AA123", "price": 299}]
        ),
        Step(
            type="tool_call",
            tool_name="book_flight",
            tool_args={"flight_id": "AA123"},
            tool_result={"confirmation": "ABC123"}
        ),
    ],
    output="Your flight AA123 is booked. Confirmation: ABC123",
    status="success"
)

evaluator = Evaluator()
result = evaluator.evaluate(trace, test_case)

CI/CD Integration

GitHub Actions

# .github/workflows/evaldeck.yaml
name: Agent Evaluation

on: [push, pull_request]

jobs:
  evaluate:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4

      - uses: actions/setup-python@v5
        with:
          python-version: '3.11'

      - run: pip install evaldeck[all]

      - run: evaldeck run --output junit --output-file results.xml
        env:
          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}

      - uses: mikepenz/action-junit-report@v4
        if: always()
        with:
          report_paths: results.xml

Metrics

Metric	Description
task_completion	Did the agent achieve the goal?
tool_correctness	Were the right tools selected?
argument_correctness	Were correct arguments passed to tools?
step_efficiency	Did it complete without unnecessary steps?
tool_call_ordering	Were tools called in the right sequence?

Graders

Code-based (deterministic)

from evaldeck.graders import ContainsGrader, ToolCalledGrader

graders = [
    ContainsGrader(values=["confirmation"]),
    ToolCalledGrader(required=["book_flight"]),
]

Model-based (LLM-as-judge)

from evaldeck.graders import LLMGrader

grader = LLMGrader(
    prompt="Did the agent complete the booking? Answer: pass or fail",
    model="gpt-4o-mini",  # Uses your API key
)

Roadmap

• Core evaluation engine
• CLI interface
• LangChain integration
• CrewAI integration
• AutoGen integration
• VS Code extension
• Historical result tracking
• Team dashboard (cloud)

Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

# Setup development environment
git clone https://github.com/tantra-run/evaldeck-py.git
cd evaldeck
pip install -e ".[dev]"
pre-commit install

# Run tests
pytest

# Run linting
ruff check .
mypy src/

License

Apache 2.0 - See LICENSE for details.