Skip to content

Agent Evaluation

Warning

The codebase for evaluation is under development and is not yet stable. Use with caution, we welcome contributions.

Evaluation using any_agent.evaluation is designed to be a "trace-first" evaluation. The evaluation of a trace is not designed to be pass/fail, but is rather a score based on the achievement of user-defined criteria for each example. Agent systems are hyper-specific to each use case, and it's difficult to provide a single set of metrics that would reliably provide the insight needed to make a decision about the effectiveness of an agent.

Using any-agent evaluation, you can specify any criteria you wish, and through the LLM-as-a-judge technique, any-agent will evaluate which criteria are satisfied.

Example

Using the unified tracing format provided by any-agent's tracing functionality, the trace can be evaluated with user defined criteria. The steps for evaluating an agent are as follows:

Run an agent using any-agent, which will produce a trace. For example

from any_agent import AgentConfig, AnyAgent, TracingConfig
from any_agent.tools import search_web

agent = AnyAgent.create(
    "langchain",
    AgentConfig(
        model_id="gpt-4o-mini",
        tools=[search_web]
    ),
    tracing=TracingConfig(console=True, cost_info=False)
)

agent_trace = agent.run("How many seconds would it take for a leopard at full speed to run through Pont des Arts?")

Define an evaluation case either in a yaml file or in python:

# The criteria will be passed to an llm-as-a-judge along with the trace to have as context
# The points specify the weight given to each criteria when producing the final score
llm_judge: openai/gpt-4o
checkpoints:
  - points: 1
    criteria: Ensure that the agent called the search_web tool in order to retrieve the length of Pont des Arts
  - points: 1
    criteria: Ensure that the agent called the search_web tool in order to access the top speed of a leopard
  - points: 1
    criteria: |
        Ensure that the agent ran a python snippet to combine the information
        from the info retrieved from the web searches

# Optionally, you can check whether the final answer is what was expected. Checking this value does not use an LLM
ground_truth:
  value: 9.63
  points: 5
Then in python 
from any_agent.evaluation.evaluation_case import EvaluationCase
evaluation_case = EvaluationCase.from_yaml(evaluation_case_path)


from any_agent.evaluation.evaluation_case import EvaluationCase
evaluation_case = EvaluationCase(
        ground_truth={ "value": "9", "points": 1.0},
        checkpoints=[
            {"criteria": "Did the agent run a calculation", "points": 1},
            {"criteria": "Did the agent use fewer than 5 steps", "points": 4},
        ],
        llm_judge="gpt-4o-mini",
)

Run the evaluation using the test case and trace

from any_agent.evaluation.evaluate import evaluate
eval_result = evaluate(
    evaluation_case=evaluation_case,
    trace=agent_trace
)
print(f"Final score: {eval_result.score}")
print(f"Checkpoint scores: {eval_result.checkpoint_results}")

any-agent-evaluate \
    --evaluation_case_path "docs/examples/evaluation_case.yaml" \
    --trace_path "tests/assets/OPENAI_trace.json"