agent-platform-eval-flywheel

Agent Platform Eval Flywheel Skill

Help users evaluate and iteratively improve GenAI models and agents using the Agent Platform GenAI Evaluation SDK (google.genai / agentplatform).

When to use this skill

• Evaluating GenAI agents or models with the Agent Platform GenAI Evaluation SDK (client.evals.evaluate()).
• Creating evaluation datasets from session traces, pandas DataFrames, or synthetic generation.
• Selecting, configuring, or writing custom evaluation metrics.
• Analyzing rubric verdicts, loss patterns, and clustering failures.
• Suggesting concrete code/prompt improvements based on eval results.
• Evaluating a model served on an Agent Platform endpoint (BYOM) or a Model-as-a-Service (MaaS) model by ID — including deploying the model first if needed. For this case, follow references/deployment.md and use the endpoint_evaluation.py / maas_evaluation.py scripts.

Safety & Confirmation Tiers (CRITICAL)

Before executing any commands or scripts on behalf of the user, you MUST adhere to the following safety tiers based on the action requested:

1. Tier R: Read-only (inspect_results.py, compare_results.py, validate_dataset.py, parse_adk_traces.py, render_html_report.py)
   • Rule: No confirmation needed. You may execute these helper scripts immediately to inspect data, validate schemas, parse traces, or compare evaluation results.
2. Tier M: Read-only with Compute Costs (client.evals.run_inference, client.evals.evaluate, client.evals.generate_user_scenarios, client.evals.generate_loss_clusters)
   • Rule: These operations invoke LLMs or remote evaluation services that consume compute resources and incur costs. This requires interactive confirmation with 'Yes'/'No' options. Once granted once, you do not have to prompt for future evaluation.

Setup

Install the SDK:

pip install google-cloud-aiplatform[evaluation]>=1.154.0 google-genai>=1.0.0

Need GOOGLE_CLOUD_PROJECT and GOOGLE_CLOUD_LOCATION. Check env vars first; if missing, ask the user. Newer Gemini models often need location="global".

The Quality Flywheel

Five stages, run in order on the first pass, then loop 2 → 5 until quality targets are met.

Shortcuts that waste time

Shortcut	Why it fails
"I'll tune the metric threshold down so it passes."	Hides real failures. Fix the agent, not the bar.
"This case is flaky, I'll skip it."	Flakiness reveals non-determinism in the agent. Fix with temperature=0 or stricter instructions.
"I just need to fix the eval dataset, not the agent."	If expected outputs keep moving, the agent has a behavior problem.
"I can tell from the trace it works — skip Stage 3."	Self-grading doesn't generalize. Always run evaluate() and read scores.
"One iteration is enough."	Expect 5–10+ iterations. Stopping early leaves regressions on other metrics undetected.

1. Prepare Data

Produce an EvaluationDataset. There are three input shapes, pick the one that matches the data the user already has:

• EvalCase list (single-turn or multi-turn):

  from agentplatform import types
  dataset = types.EvaluationDataset(eval_cases=[
      types.EvalCase(prompt="What is 2+2?", response="4", reference="4"),
      # For multi-turn agent traces, set agent_data instead of prompt/response.
  ])
  

  Multi-turn agent traces wrap each conversation in AgentData → ConversationTurn → AgentEvent. See references/dataset_schema.md for the full type hierarchy.

• Pandas DataFrame (tabular sources — CSV, BigQuery, Sheets):

  import pandas as pd
  from agentplatform import types
  
  df = pd.DataFrame({
      "prompt":    ["What is 2+2?", "Capital of France?"],
      "response":  ["4",            "Paris"],
      "reference": ["4",            "Paris"],
  })
  dataset = types.EvaluationDataset(eval_dataset_df=df)
  

  Column names must match the fields the chosen metrics expect (see references/dataset_schema.md for the per-metric requirements table).

• Cold start (no data at all): synthesize scenarios server-side with client.evals.generate_user_scenarios(...) and a UserScenarioGenerationConfig (user_scenario_count, simulation_instruction, environment_data). Stage 2 plays them out.

For ADK session dumps, use scripts/parse_adk_traces.py instead of writing the conversion by hand.

2. Run Inference

Populate responses/traces on the dataset. Skip this stage if traces are already complete (e.g., production logs or replay).

# Agent eval — pass a callable wrapping the user's ADK Agent/App.
client.evals.run_inference(model=agent_callable, src=dataset)

# Model eval — pass a model ID directly.
client.evals.run_inference(model="gemini-2.5-flash", src=dataset)

# Synthesized scenarios — let the simulator drive.
client.evals.run_inference(
    model=agent_callable,
    src=dataset,
    user_simulator_config=UserSimulatorConfig(max_turn=10),
)

# DataFrame also works as src= — no EvalCase wrapping needed.
client.evals.run_inference(model="gemini-2.5-flash", src=df)

3. Grade (always run)

result = client.evals.evaluate(dataset=dataset, metrics=[...])

Pick metrics by what you want to measure. Full catalog in references/metric_registry.md.

Agent metrics (multi-turn, adaptive rubrics) — start here for agent eval.

Goal	Metric
Did the agent achieve the user's goal?	multi_turn_task_success
Was the reasoning path logical and efficient?	multi_turn_trajectory_quality
Tool/function calling quality across turns	multi_turn_tool_use_quality
Overall conversational quality	multi_turn_general_quality
Final response quality (no reference needed)	final_response_quality
Final response vs. a golden reference	final_response_match
Single-turn tool use	tool_use_quality

General quality metrics (single-turn, adaptive rubrics) — for model eval.

Goal	Metric
Overall response quality (recommended starting point)	general_quality
Linguistic quality (fluency, coherence, grammar)	text_quality
Adherence to specific constraints / instructions	instruction_following

Static rubric metrics (fixed criteria) — apply alongside the above.

Goal	Metric
Catch hallucinated claims (RAG, factual answers)	hallucination
Factuality / consistency against provided context	grounding
Safety policy compliance	safety

Domain-specific check no built-in covers: write a custom metric.

• Predefined: types.RubricMetric.<NAME> — server-side AutoRater, no judge model needed.
• Custom LLM-as-a-judge: types.LLMMetric with prompt_template or types.MetricPromptBuilder for structured rubrics.
• Custom code: types.CodeExecutionMetric with a custom_function string containing def evaluate(instance: dict) for remote sandboxed execution; or types.Metric with custom_function=<callable> for local execution.

Always persist the result so Stage 4 and 5 can read it. Save both JSON (machine-readable, diffable) and HTML (human-readable, linkable):

import datetime
from pathlib import Path

from agentplatform._genai import _evals_visualization

out_dir = Path("artifacts/grade_results")
out_dir.mkdir(parents=True, exist_ok=True)
ts = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")

result_json = result.model_dump_json()
(out_dir / f"results_{ts}.json").write_text(result_json)

html = _evals_visualization.get_evaluation_html(result_json)
(out_dir / f"results_{ts}.html").write_text(str(html))

Or after the fact: scripts/render_html_report.py --type evaluation or scripts/inspect_results.py --save-html.

4. Analyze Failures

Read summary_metrics and eval_case_results — never fabricate scores. Use scripts/inspect_results.py --failing-only to filter to failures.

For each failed metric, see references/failure_patterns.md for deeper diagnoses. The compact mapping:

Failing metric	What to change
multi_turn_task_success low	The agent isn't completing the goal — fix orchestration, missing tool calls, premature termination, wrong tool selection.
multi_turn_trajectory_quality low	The agent reaches the goal inefficiently — refine planning prompts, remove redundant tool calls.
multi_turn_tool_use_quality low	Fix tool descriptions, parameter docstrings, or agent instructions for tool selection.
final_response_quality low	Read auto-generated rubric verdicts; refine instructions to address the worst-scoring criterion.
final_response_match low	The agent's final answer doesn't match the golden reference — adjust response format or update the reference.
hallucination low	Tighten instructions to stay grounded in tool output; verify the tool actually returned the claimed data.
grounding low	The response contradicts the provided context — add explicit "cite only from context" instructions.
safety low	Add safety guardrails; review the violating content category in the rubric verdict.
general_quality / text_quality low	Adjust system instruction wording; the model's default phrasing is too generic for the task.
instruction_following low	The agent is ignoring constraints — restate them in the system instruction or use stricter wording.
Agent calls wrong tools	Fix tool descriptions, agent instructions, or tool_config.
Agent calls extra tools	Add explicit stop instructions, or switch to multi_turn_tool_use_quality to surface the extra calls in the rubric.

For 10+ failures on the same metric, use the Error Analysis service to cluster failures into themes (L1/L2 taxonomy categories) instead of reading every trace:

# Only supports multi_turn_task_success and multi_turn_tool_use_quality.
# Service runs in the global region.
analysis_client = agentplatform.Client(project="PROJECT_ID", location="global")
response = analysis_client.evals.generate_loss_clusters(
    eval_result=result,
    metric="multi_turn_task_success",
    config={"max_top_cluster_count": 5},
)
for r in response.results:
    for cluster in r.clusters:
        print(
            f"[{cluster.taxonomy_entry.l1_category}/"
            f"{cluster.taxonomy_entry.l2_category}] "
            f"{cluster.item_count} cases — {cluster.taxonomy_entry.description}"
        )

Save response.model_dump_json() and render with scripts/render_html_report.py --type loss-analysis.

5. Optimize & Iterate

Apply a fix targeting the failing metric. Re-run Stage 3. Compare with scripts/compare_results.py --baseline <prev> --candidate <new> to confirm the target improved AND no other metric regressed.

Track progress across iterations:

Iteration	Metric A	Metric B	Change made
Baseline	0.62	0.55	—
v2	0.78	0.68	Added grounding prompt
v3	0.81	0.72	Fixed tool selection

Expect 5–10+ iterations per failing case. Only after a case passes should you expand coverage with more eval cases.

Proving your work

Never claim eval results you didn't read from an actual result object.

• After running eval, print the summary_metrics table (scripts/inspect_results.py).
• After a fix, show before/after via scripts/compare_results.py.
• Before declaring success, confirm ALL cases pass — not just the one you were working on.

If you can't produce the evidence (SDK call failed, result truncated, metric unsupported), say so explicitly. Don't paper over gaps.

Rules of Engagement

1. Always Plan First: Before writing a script, output a <plan> block detailing the steps you are about to take.
2. Step-by-Step Execution: Write the script, execute it, wait for output, then analyze. Don't do everything in one response.
3. Standard Python: Use standard Python imports (import agentplatform, from google.genai import types). Don't use internal import paths.
4. Verify Before Guessing: When unsure about SDK types or metrics, check the SDK source code rather than guessing or hallucinating.

SDK Quick Reference

import agentplatform
from agentplatform import types
from google.genai import types as genai_types
import pandas as pd

# Initialize client
client = agentplatform.Client(project="PROJECT_ID", location="LOCATION")

# --- SINGLE-TURN EVAL (EvalCase list) ---
dataset = types.EvaluationDataset(eval_cases=[
    types.EvalCase(prompt="Query here", response="Model response here"),
])

# --- SINGLE-TURN EVAL (pandas DataFrame) ---
df = pd.DataFrame({
    "prompt":   ["Q1", "Q2"],
    "response": ["A1", "A2"],
})
dataset = types.EvaluationDataset(

...