AgentAsJudgeEval

AgentAsJudgeEval grades an input/output pair against custom criteria with an LLM judge. Score recorded outputs directly with run(), or attach the eval as a post_hook to grade live runs.

from agno.eval import AgentAsJudgeEval, AgentAsJudgeResult

Parameters

Parameter	Type	Default	Description
`criteria`	`str`	`""`	Criteria the judge grades the output against.
`scoring_strategy`	`Literal["numeric", "binary"]`	`"binary"`	How the judge grades. `"binary"` returns a pass/fail verdict. `"numeric"` returns a 1-10 score that passes when it meets `threshold`.
`threshold`	`int`	`7`	Minimum passing score from 1 to 10. Only used when `scoring_strategy` is `"numeric"`.
`on_fail`	`Optional[Callable[[AgentAsJudgeEvaluation], None]]`	`None`	Callback invoked with the evaluation when it fails. Async callbacks require `arun()`.
`additional_guidelines`	`Optional[Union[str, List[str]]]`	`None`	Guidelines appended to the default judge’s instructions.
`name`	`Optional[str]`	`None`	Evaluation name.
`model`	`Optional[Model]`	`None`	Model for the default judge agent. Falls back to OpenAI `gpt-5-mini` when unset.
`evaluator_agent`	`Optional[Agent]`	`None`	Custom judge agent. Its `output_schema` is replaced with the response schema matching `scoring_strategy`.
`print_summary`	`bool`	`False`	Print the summary table with pass rate and score statistics after the run.
`print_results`	`bool`	`False`	Print the per-evaluation results after the run.
`show_spinner`	`bool`	`True`	Render the progress spinner. Set to `False` when the eval runs inside code that must not write to the console.
`file_path_to_save_results`	`Optional[str]`	`None`	File path where the result is saved.
`debug_mode`	`bool`	`False`	Enable debug logs. Defaults to `True` when the `AGNO_DEBUG` environment variable is `"true"`.
`db`	`Optional[Union[BaseDb, AsyncBaseDb]]`	`None`	Database where the eval run is stored.
`telemetry`	`bool`	`True`	Log minimal telemetry for the eval run.
`run_in_background`	`bool`	`False`	Run the eval in the background when attached as a `post_hook`, without blocking the agent’s response. Takes effect only when the run is served through AgentOS, which supplies the background task queue; direct `run()` and `arun()` calls on the agent execute the eval inline.

Methods

`run()` and `arun()`

Both accept the same keyword arguments and return Optional[AgentAsJudgeResult]. Provide either input and output for a single evaluation, or cases for batch evaluation. run() raises with an async database; use arun() instead.

Argument	Type	Default	Description
`input`	`Optional[str]`	`None`	Input text for a single evaluation.
`output`	`Optional[str]`	`None`	Output text for a single evaluation.
`cases`	`Optional[List[Dict[str, str]]]`	`None`	Batch of dicts with `input` and `output` keys, evaluated in order.
`print_summary`	`bool`	`False`	Print the summary table after the run.
`print_results`	`bool`	`False`	Print the per-evaluation results after the run.

Hook methods

AgentAsJudgeEval extends BaseEval, so an instance can be passed directly to an Agent’s or Team’s post_hooks. The post_check() and async_post_check() methods grade the run’s input and output and log the result to db with the agent or team ID attached. Pre-hooks are not supported and raise ValueError.

AgentAsJudgeResult

Field	Type	Description
`run_id`	`str`	Unique ID for this evaluation run.
`results`	`List[AgentAsJudgeEvaluation]`	One entry per evaluated input/output pair.
`avg_score`	`Optional[float]`	Average score. `None` in binary mode.
`min_score`	`Optional[float]`	Lowest score. `None` in binary mode.
`max_score`	`Optional[float]`	Highest score. `None` in binary mode.
`std_dev_score`	`Optional[float]`	Standard deviation of scores. `None` in binary mode.
`pass_rate`	`float`	Percentage of evaluations that passed, from 0 to 100.

AgentAsJudgeResult also exposes print_summary(console=None) and print_results(console=None).