Evals

Lock in agent behavior with regression tests.

Next, let’s lock in our agent behavior with evals.

Think of Evals as regression tests for your agents. Same prompts, same agents, run on a schedule. Notify when behavior drifts.

When we run docs/improve-agent.md, we’re looking for out-of-distribution improvements. Evals make sure in-distribution cases continue to pass. The two work together.

Cases

Cases live in evals/cases.py. Each case sends one input to an agent and (optionally) checks two things:

• judge: AgentAsJudgeEval scores the response against criteria (binary pass/fail) using an LLM.
• reliability: ReliabilityEval checks which tools fired against expected_tool_calls.

Results are stored in your database via eval_db (visible at os.agno.com).

A case looks like this:

evals/cases.py

CASES: tuple[Case, ...] = (
    Case(
        name="web_search_recent_anthropic_research",
        agent=web_search,
        input="What did Anthropic publish about agents recently?",
        criteria=(
            "Answers the question by citing at least one real Anthropic URL "
            "(anthropic.com domain). The response is grounded in fetched content."
        ),
        expected_tool_calls=(_WEB_SEARCH_TOOL,),
    ),
    # add more cases here
)

A case can use either check or both. If both are set, the agent runs once and feeds the same response into both.

Run the suite

1. Create a virtual environment

   To run the eval suite, let’s create a local virtual environment

   ./scripts/venv_setup.sh
   

   Activate it

   source .venv/bin/activate
   
2. Run the eval suite

   python -m evals                # full suite
   

   Other options:

   python -m evals -v             # stream the agent run with full panels
   python -m evals --case <name>  # single case while iterating
   

Each case prints the response, the judge verdict, and the reliability verdict. The run ends with an Eval Summary table.

Results write to Postgres via eval_db. You can view the Eval history on os.agno.com alongside your sessions and traces. You can see when a case started failing and what changed.

Diagnose failures with Claude Code

Open Claude Code and paste:

Run docs/eval-and-improve.md

Claude runs the full suite, triages every failure (bad criteria, real regression, flaky LLM judge), and proposes in-scope fixes. It edits the agent or the case, re-runs, and shows you the diff.

When to run evals

Trigger	Frequency
Before deploying a change to an agent	Every time
As part of CI	Every PR
Against production	On a weekly cron
After bumping a model version	Every time

The weekly production cron is the most valuable one. Wire it into your platform’s scheduler. See scheduling for the cron API.

What good cases look like

• Specific. “Returns a JSON object with ticker and price” beats “Returns the right answer”.
• Stable. Avoid prompts whose correct answer changes daily. Use phrasing like “describes a real, recent…” instead of locking in a specific result.
• Scoped to one behavior. One case per behavior makes failures easy to read.
• Anchored to tools. expected_tool_calls catches the failure mode where the agent confidently makes things up instead of calling a tool.

Next

Next steps →