Agent Evaluation

Agent Evaluation with MLflow

Comprehensive guide for evaluating GenAI agents with MLflow. Use this skill for the complete evaluation workflow or individual components - tracing setup, environment configuration, dataset creation, scorer definition, or evaluation execution. Each section can be used independently based on your needs.

⛔ CRITICAL: Must Use MLflow APIs

DO NOT create custom evaluation frameworks. You MUST use MLflow's native APIs:

• Datasets: Use mlflow.genai.datasets.create_dataset() - NOT custom test case files
• Scorers: Use mlflow.genai.scorers and mlflow.genai.judges.make_judge() - NOT custom scorer functions
• Evaluation: Use mlflow.genai.evaluate() - NOT custom evaluation loops
• Scripts: Use the provided scripts/ directory templates - NOT custom evaluation/ directories

Why? MLflow tracks everything (datasets, scorers, traces, results) in the experiment. Custom frameworks bypass this and lose all observability.

If you're tempted to create evaluation/eval_dataset.py or similar custom files, STOP. Use scripts/create_dataset_template.py instead.

Table of Contents

1. Quick Start
2. Documentation Access Protocol
3. Setup Overview
4. Evaluation Workflow
5. References

Quick Start

⚠️ REMINDER: Use MLflow APIs from this skill. Do not create custom evaluation frameworks.

Setup (prerequisite): Install MLflow 3.8+, configure environment, integrate tracing

Evaluation workflow in 5 steps (each uses MLflow APIs):

1. Understand: Run agent, inspect traces, understand purpose
2. Scorers: Register the agreed scorer suite (use build-a-scorer first if it does not exist yet)
3. Dataset: ALWAYS discover existing datasets first, only create new if needed 3.5. Dry Run: Run 3 questions first — catch broken tools and misconfigured scorers before full eval
4. Evaluate: Run agent on dataset, apply scorers, analyze results

Scope boundary: this skill owns running evaluation — registration, datasets, execution, and analysis. Designing the scorer suite (which criteria, and code check vs. built-in vs. LLM judge) is the build-a-scorer skill's job. The seam is a confirmed suite of criteria with an implementation chosen for each.

Command Conventions

Always use uv run for MLflow and Python commands:

uv run mlflow --version          # MLflow CLI commands
uv run python scripts/xxx.py     # Python script execution
uv run python -c "..."           # Python one-liners

This ensures commands run in the correct environment with proper dependencies.

CRITICAL: Separate stderr from stdout when capturing CLI output:

When saving CLI command output to files for parsing (JSON, CSV, etc.), always redirect stderr separately to avoid mixing logs with structured data:

# Save both separately for debugging
uv run mlflow traces evaluate ... --output json > results.json 2> evaluation.log

Documentation Access Protocol

All MLflow documentation must be accessed through llms.txt:

1. Start at: https://mlflow.org/docs/latest/llms.txt
2. Query llms.txt for your topic with specific prompt
3. If llms.txt references another doc, use WebFetch with that URL
4. Do not use WebSearch - use WebFetch with llms.txt first

This applies to all steps, especially:

• Dataset creation (read GenAI dataset docs from llms.txt)
• Scorer registration (check MLflow docs for scorer APIs)
• Evaluation execution (understand mlflow.genai.evaluate API)

Discovering Agent Structure

Each project has unique structure. Use dynamic exploration instead of assumptions:

Find Agent Entry Points

# Search for main agent functions
grep -r "def.*agent" . --include="*.py"
grep -r "def (run|stream|handle|process)" . --include="*.py"

# Check common locations
ls main.py app.py src/*/agent.py 2>/dev/null

# Look for API routes
grep -r "@app\.(get|post)" . --include="*.py"  # FastAPI/Flask
grep -r "def.*route" . --include="*.py"

Understand Project Structure

# Check entry points in package config
cat pyproject.toml setup.py 2>/dev/null | grep -A 5 "scripts\|entry_points"

# Read project documentation
cat README.md docs/*.md 2>/dev/null | head -100

# Explore main directories
ls -la src/ app/ agent/ 2>/dev/null

Setup Overview

Pre-check: Use Existing Environment

Before doing ANY setup, check if MLFLOW_TRACKING_URI and MLFLOW_EXPERIMENT_ID are already set:

echo "MLFLOW_TRACKING_URI=$MLFLOW_TRACKING_URI"
echo "MLFLOW_EXPERIMENT_ID=$MLFLOW_EXPERIMENT_ID"

If BOTH are already set, skip Steps 1-2 entirely. The environment is pre-configured. Do NOT run setup_mlflow.py, do NOT create a .env file, do NOT override these values. Go directly to Step 3 (tracing integration) and the evaluation workflow.

Setup Steps (only if environment is NOT pre-configured)

1. Install MLflow (version >=3.8.0)
2. Configure environment (tracking URI and experiment)
   • Guide: Follow references/setup-guide.md Steps 1-2
3. Integrate tracing (autolog and @mlflow.trace decorators)
   • ⚠️ MANDATORY: Use the instrumenting-with-mlflow-tracing skill for tracing setup
   • ✓ VERIFY: Run scripts/validate_tracing_runtime.py after implementing

⚠️ Tracing must work before evaluation. If tracing fails, stop and troubleshoot.

Checkpoint - verify before proceeding:

• MLflow >=3.8.0 installed
• MLFLOW_TRACKING_URI and MLFLOW_EXPERIMENT_ID set
• Autolog enabled and @mlflow.trace decorators added
• Test run creates a trace (verify trace ID is not None)

Validation scripts:

uv run python scripts/validate_environment.py  # Check MLflow install, env vars, connectivity
uv run python scripts/validate_auth.py         # Test authentication before expensive operations

Evaluation Workflow

Step 1: Agent Interview (REQUIRED — do not skip)

Before doing anything else, ask the user these questions. Do NOT proceed until you have answers.

Required:

1. "What does your agent do? Describe its purpose in 1-2 sentences."
2. "What are the 2-3 most important things it needs to get right?"
3. "Are there common failure modes you've already noticed?"

Use answers to:

• Write the agent_description parameter for generate_evals_df
• Set evaluation priorities
• Decide whether the quality criteria are already clear (continue to Step 2) or still need to be worked out with the user (use the build-a-scorer skill — see Step 2)

If running in automated mode: Read agent purpose from the codebase (SKILL.md, README, or main entry point docstring). Still surface what you found and confirm before proceeding.

Scope note: these three questions establish what the agent is. They are not enough to design a scorer suite. If the answers to (2) and (3) are vague — "good answers", "it should be accurate" — do not invent scorer names from them; that is what build-a-scorer is for.

Step 2: Register Quality Scorers

This step registers and runs scorers. Designing them — which criteria matter, and whether each is cheapest as a code check, a built-in, or an LLM judge — belongs to the build-a-scorer skill.

💡 If the user does not already have a scorer suite, use the build-a-scorer skill first. It returns a confirmed set of 3-5 atomic criteria, each with an implementation chosen per criterion (code check, built-in, or make_judge) and an output type that survives aggregation. Come back here to register and run them.

Go straight to registration below when scorers are already registered in the experiment, or the user hands you an explicit list of criteria and implementations.

1. Check registered scorers in your experiment:

   uv run mlflow scorers list -x $MLFLOW_EXPERIMENT_ID
   

IMPORTANT: if there are registered scorers in the experiment then they must be used for evaluation.

2. Add the built-in scorers from the agreed suite

See references/scorers.md for the built-in scorers. Add any that the suite calls for and that are not already registered.

3. Add the custom scorers from the agreed suite

Implement the remaining criteria with make_judge() (LLM judges) or @scorer (deterministic code checks). See references/scorers.md for the APIs and references/scorers-constraints.md for constraints that silently break results.

If you are inventing criteria at this point rather than implementing agreed ones, stop and use build-a-scorer — a scorer whose criterion the user has not confirmed will not survive validation.

⚠️ CRITICAL — Scorer Return Values: Scorers MUST instruct the LLM judge to return "yes" or "no" (or booleans/numerics). Return values of "pass" or "fail" are silently cast to None by _cast_assessment_value_to_float and excluded from results.metrics with no error or warning — results simply disappear. See references/scorers-constraints.md Constraint 2 for the full list of safe vs. broken return values.

4. REQUIRED: Register new scorers before evaluation using Python API:

   from mlflow.genai.judges import make_judge
   from mlflow.genai.scorers import BuiltinScorerName
   import os
   
   scorer = make_judge(...)  # Or, scorer = BuiltinScorerName()
   scorer.register()
   

** IMPORTANT: See references/scorers.md → "Model Selection for Scorers" to configure the model parameter of scorers before registration.

⚠️ Scorers MUST be registered before evaluation. Inline scorers that aren't registered won't appear in mlflow scorers list and won't be reusable.

5. Verify registration:

   uv run mlflow scorers list -x $MLFLOW_EXPERIMENT_ID  # Should show your scorers
   

Step 3: Prepare Evaluation Dataset

ALWAYS discover existing datasets first to prevent duplicate work:

1. Run dataset discovery (mandatory):

   uv run python scripts/list_datasets.py  # Lists, compares, recommends datasets
   uv run python scripts/list_datasets.py --format json  # Machine-readable output
   uv run python scripts/list_datasets.py --help  # All options
   

2. Present findings to user:

   • Show all discovered datasets with their characteristics (size, topics covered)
   • If datasets found, highlight most relevant options based on agent type

3. Ask user about existing datasets:

   • "I found [N] existing evaluation dataset(s). Do you want to use one of these? (y/n)"
   • If yes: Ask which dataset to use and record the dataset name — skip to Step 3.5
   • If no: Proceed to Phase A below

If creating a new dataset, use the two-phase approach below.

Phase A: Sanity Check (5 questions — always run first)

Create a minimal 5-question dataset manually from the Step 1 interview answers. The goal is to confirm the pipeline works end-to-end before investing in large-scale generation.

import mlflow
from mlflow.genai.datasets import create_dataset

# Derive 5 representative questions directly from the agent's stated purpose
# and known failure modes identified in Step 1
sanity_records = [
    {"inputs": {"query": "<question 1 from interview>"}, "expected_response": "<expected answer>"},
    {"inputs": {"query": "<question 2 from interview>"}, "expected_response": "<expected answer>"},
    # ... 5 total
]

sanity_dataset = create_dataset(
    records=sanity_records,
    name="sanity-check-5q",
)

Run evaluation on this dataset (see Step 4), then present results to the user with this framing:

"This is a sanity check — 5 questions confirm the pipeline works but aren't statistically meaningful. Proceeding to Phase B to generate a proper evaluation set."

Only proceed to Phase B once Phase A completes without errors.

Phase B: Proper Evaluation Dataset (100+ questions — run after Phase A passes)

Generate questions from the agent's actual corpus rather than inventing them from scratch. The approach depends on whether the project uses Databricks or OSS MLflow.

On Databricks — use generate_evals_df to synthesize questions from the agent's document corpus:

from databricks.agents.evals import generate_evals_df, estimate_synthetic_num_evals
import mlflow

# agent_description comes from Step 1 interview answers
agent_description = "<agent purpose from interview>"

# docs_df: a Spark or pandas DataFrame with a "content" column containing
# the documents/chunks the agent retrieves from (e.g., your Vector Search index)
evals = generate_evals_df(
    docs=docs_df,
    num_evals=100,
    agent_description=agent_description,
)

# Merge into MLflow dataset — don't create a separate dataset
dataset = mlflow.genai.datasets.create_dataset(name="generated-evals-100q")
dataset.merge_records(evals)

To estimate the right num_evals before generating:

recommended = estimate_synthetic_num_evals(docs_df)
print(f"Recommended num_evals: {recommended}")

Dataset size guidance:

• <30 questions: not statistically meaningful — avoid drawing conclusions
• 50–100 questions: adequate for catching regressions, suitable for most agents
• 200+ questions: recommended when comparing model variants or scoring multiple dimensions

On OSS MLflow — use RAGAS TestsetGenerator to generate from your document corpus:

from ragas.testset import TestsetGenerator
from ragas.llms import LangchainLLMWrapper
from ragas.embeddings import LangchainEmbeddingsWrapper

generator = TestsetGenerator(
    llm=LangchainLLMWrapper(your_llm),
    embedding_model=LangchainEmbeddingsWrapper(your_embeddings),
)
testset = generator.generate_with_langchain_docs(docs, testset_size=100)
evals_df = testset.to_pandas()

# Convert to MLflow dataset schema and merge
import mlflow
records = [
    {"inputs": {"query": row["user_input"]}, "expected_response": row["reference"]}
    for _, row in evals_df.iterrows()
]
dataset = mlflow.genai.datasets.create_dataset(name="generated-evals-100q")
dataset.merge_records(records)

If no document corpus is available — ask the user to provide 50–100 representative queries from production logs or usage history. These are more realistic than synthetic questions and are preferable when available.

IMPORTANT: Do not skip dataset discovery. Always run list_datasets.py first, even if you plan to create a new dataset. This prevents duplicate work and ensures users are aware of existing evaluation datasets.

For complete dataset guide: See references/dataset-preparation.md

Checkpoint - verify before proceeding:

• Scorers have been registered
• Phase A sanity check passed (pipeline runs end-to-end)
• Phase B dataset created with 50+ questions (or existing dataset selected)

Step 3.5: Dry Run (REQUIRED before full eval)

Run evaluation on 3 questions from the dataset before committing to the full run. This catches broken tools, misconfigured scorers, and auth failures early — before they silently corrupt 100-question results.

If you completed Phase A above, the pipeline is already validated — focus the dry run on scorer output only.

import mlflow

dataset = mlflow.genai.datasets.get_dataset(name="<your-dataset-name>")
dry_run_records = dataset.df.head(3)

Run mlflow.genai.evaluate() on these 3 records using the same wrapper and scorers as the full eval.

For each response, check:

1. Tool calls — Did the agent call any tools? If it called zero tools on questions that require retrieval, tools are likely broken (403s, rate limits, missing credentials).
2. Response quality — Are responses empty or generic ("I don't know", "I can't help with that")? Empty responses score as irrelevant and will skew the full eval.
3. Scorer output — Did all 3 scores come back as 0 or None? If so, the scorer is misconfigured (check return values — "pass"/"fail" are silently cast to None; use "yes"/"no" instead).

Decision gate:

• If dry run shows tool failures or empty responses: Stop. Fix the underlying issue (auth, tool config, retrieval) before proceeding. Do not run the full eval on broken infrastructure.
• If all 3 scorer outputs are 0 or None: Stop. Debug scorer return values and re-register before proceeding.
• If dry run passes: Report to the user: "Dry run passed (3/3 responses non-empty, tools called, scores non-zero). Proceeding to full eval." Then continue to Step 4.

Why this matters: Tool failures (403s from docs scraping, GitHub API rate limits) produce empty agent responses that score as 0. Running a 100-question eval only to discover all tools were failing wastes time and produces misleading results. The dry run catches this in under a minute.

Step 4: Run Evaluation

Large datasets (50+ questions)? See references/throughput-guide.md for throughput optimization — covers parallelism env vars, async predict functions, and dataset splitting for 200+ question evals.

4a. Estimate Runtime Before Starting

Before launching evaluation, tell the user how long it will take:

1. Count the dataset questions:

   import mlflow
   dataset = mlflow.genai.datasets.get_dataset(name="<your-dataset-name>")
   print(f"Dataset size: {len(dataset.df)} questions")
   

2. Calculate the estimate — each question runs the agent once and the judge scorer once:

   • Opus-class judge models (e.g. claude-opus-4): ~45–90s per question
   • Sonnet-class judge models (e.g. claude-sonnet-4): ~20–45s per question
   • Multiple scorers per question add time proportionally

   Estimated time = N questions × 30–60s per question ÷ parallelism factor (typically 4–8x)
   

3. Tell the user before starting:

   "This dataset has N questions. At ~30–60s per question with typical parallelism, evaluation will take approximately X–Y minutes. I'll run it as a background task so you can continue working — I'll summarize the results when it's done."

4b. Generate the Evaluation Script

# Generate evaluation script (specify module and entry point)
uv run python scripts/run_evaluation_template.py \
  --module mlflow_agent.agent \
  --entry-point run_agent

The generated script creates a wrapper function that:

• Accepts keyword arguments matching the dataset's input keys
• Provides any additional arguments the agent needs (like llm_provider)
• Runs mlflow.genai.evaluate(data=df, predict_fn=wrapper, scorers=registered_scorers)
• Saves results to evaluation_results.csv

⚠️ CRITICAL: wrapper Signature Must Match Dataset Input Keys

MLflow calls predict_fn(**inputs) - it unpacks the inputs dict as keyword arguments.

Common Mistake (WRONG):

def wrapper(inputs):  # ❌ WRONG - inputs is NOT a dict
    return agent(inputs["query"])

4c. Launch as a Background Sub-Agent

Run the evaluation as a background sub-agent so the main session stays available. Use the Agent tool with run_in_background: true:

Sub-agent instructions (pass these verbatim):

Run the agent evaluation and write results to scratchpad.

Steps:
1. cd <project-directory>
2. Run: uv run python run_agent_evaluation.py
3. When complete, write a summary to scratchpad/eval-results.md with:
   - Exit status (success or error message)
   - Path to results file (evaluation_results.csv)
   - Wall-clock time taken
4. Return only: "Evaluation complete. Results written to scratchpad/eval-results.md"

In the main session, poll for completion by checking for the scratchpad file rather than blocking:

# Poll every 30s using Glob
# Glob("scratchpad/eval-results.md")
# When the file appears, read it and proceed to analysis

Do NOT use TaskOutput to wait for the background agent — that dumps the full transcript (~10–20k tokens) into the main context.

4d. Analyze Results (after evaluation completes)

Once scratchpad/eval-results.md appears, run analysis:

# Pattern detection, failure analysis, recommendations
# Reads the CSV produced by mlflow.genai.evaluate() above
uv run python scripts/analyze_results.py evaluation_results.csv

Generates evaluation_report.md with per-scorer pass rates and improvement suggestions.

The script reads {scorer_name}/value and {scorer_name}/rationale columns from the CSV. It also accepts the legacy JSON format from mlflow traces evaluate for backward compatibility:

uv run python scripts/analyze_results.py evaluation_results.json  # legacy format
uv run python scripts/analyze_results.py evaluation_results.csv --output my_report.md  # custom output

References

Detailed guides in references/ (load as needed):

• setup-guide.md - Environment setup (MLflow install, tracking URI configuration)
• Tracing: Use the instrumenting-with-mlflow-tracing skill (authoritative guide for autolog, decorators, session tracking, verification)
• dataset-preparation.md - Dataset schema, APIs, creation, Unity Catalog
• scorers.md - Built-in vs custom scorers, registration, testing
• scorers-constraints.md - CLI requirements for custom scorers (yes/no format, templates)
• troubleshooting.md - Common errors by phase with solutions
• throughput-guide.md - Parallelism env vars, async predict_fn, dataset splitting for 200+ question evals

Scripts are self-documenting - run with --help for usage details.