Agent Evaluation Guide

Requires: agents-cli (uv tool install google-agents-cli) — install uv first if needed.

Scaffolded project? If you used /google-agents-cli-scaffold, you already have agents-cli eval run (chains generate + grade), tests/eval/datasets/, and tests/eval/eval_config.yaml. Start with executing eval run and iterate from there.

Reference Files

The Quality Flywheel

Improving agent quality is iterative. The 4 stages below describe the loop. Each stage has a Default path (you, the coding agent, do the work directly) and an Opt-in CLI command that delegates to the Agent Platform Eval Service for better quality and scale.

1. Prepare Data

Default: Use or edit the scaffolded tests/eval/datasets/basic-dataset.json to define single-turn eval inputs. Start with 1–2 cases.

Opt-in: agents-cli eval dataset synthesize: user-simulate multi-turn datasets when you lack data; its output already includes traces, so Stage 2 collapses to agents-cli eval grade alone. See Eval Commands and references/user-simulation.md.

2. Run the Eval (always run)

Default: agents-cli eval run runs the agent over the dataset and grades the traces, writing results_<ts>.{json,html} to artifacts/grade_results/.

Decoupled form: eval generate then eval grade, for a custom traces location, re-grading without re-running the agent, or traces from synthesize (eval grade alone).

3. Analyze Failures

Default: Open the latest artifacts/grade_results/results_<ts>.html (or .json) and identify failed metrics — see What to fix when scores fail below for the fix table.

Opt-in: agents-cli eval analyze, LLM-based failure clustering; prefer when you have 10+ failing cases and want categorized failure modes. See references/advanced-commands.md.

4. Optimize & Code Fix

Default: Edit the agent — adjust prompts, tool descriptions, instructions, or eval dataset based on the failure analysis. See What to fix when scores fail below for the failure → fix mapping.

Opt-in: agents-cli eval optimize runs ADK GEPA prompt optimization against a target metric (see references/advanced-commands.md). Suitable for prompt-only failures. The optimized prompt appears in the command output; capture it and apply it to the agent. For the full per-iteration trace, set print_detailed_results: true in your optimization config file.

Long-running and expensive. GEPA optimization makes many LLM calls and can take a long time. Do not run it unless the user explicitly asks for prompt optimization. When you do run it, iterate as far as possible with manual fixes first, then run a single final eval optimize — never loop on this command.

Running the loop

Iterate stages 2 → 3 → 4 → 2 (with synthesize, re-run Stage 1 each pass, then eval grade). After each fix, run agents-cli eval compare <prev_results>.json <new_results>.json to confirm the target metric improved without regressing others. Expect 5–10+ iterations per case before it passes, which is normal. Only after a case passes should you expand coverage with more eval cases.

When doing 5+ iterations, maintain a task list of which cases are fixed, which are still failing, and what fixes you've tried. Prevents re-attempting the same fix.

Hold cases back. Keep a slice of cases out of the loop and grade them only when you think you're done — otherwise you can't tell a fix that generalizes from one fitted to the cases you iterated against.

Shortcuts That Waste Time

Recognize these rationalizations and push back — they always cost more time than they save:

Choosing the Right Metrics

Pick built-in metrics by what you want to measure. Only multi_turn_task_success, multi_turn_trajectory_quality, and multi_turn_tool_use_quality accept multi-turn traces; every other built-in 400s on one. When no built-in fits, write a custom metric (see Evaluation Configuration Schema below).

Run agents-cli eval metric list to see all available built-ins. For full metric definitions and rubric details, see the Agent Platform metric docs and references/metrics-guide.md.

What to fix when scores fail

After agents-cli eval run completes, inspect the latest artifacts/grade_results/results_<timestamp>.json (or open the .html file) for per-case scores and judge rationales, the input to every fix decision below.

After applying a fix, rerun agents-cli eval run and use agents-cli eval compare <prev_results>.json <new_results>.json to confirm the fix improved the target metric without regressing others.

Eval Commands

agents-cli eval <subcommand> --help is the authoritative flag list; the examples below are the common invocations.

eval run (default)

Runs the agent over the dataset and grades the traces in one command.

# Basic: dataset from tests/eval/datasets/, results to artifacts/grade_results/,
# metrics from tests/eval/eval_config.yaml
agents-cli eval run

# Advanced: pick the dataset, metrics, and output dir
agents-cli eval run --dataset tests/eval/datasets/custom.json --metrics final_response_quality,safety --output ./out/

eval generate

Runs an agent over an evaluation dataset and writes traces to disk.

By default, runs the agent in a local HTTP server (launches the project's fast_api_app.py if it exists, or falls back to adk api_server) and sends each evaluation case over HTTP. You can generate traces from an already-running agent by passing its HTTP endpoint and app name to --url and --app-name.

# Basic — uses tests/eval/datasets/, writes to artifacts/traces/
agents-cli eval generate

# Advanced — custom dataset and output dir
agents-cli eval generate --dataset tests/eval/datasets/custom.json -o ./custom_traces/

# Against a deployed agent (or one you started manually)
agents-cli eval generate --url https://my-agent.run.app --app-name app

eval grade

Scores traces (from eval generate, eval dataset synthesize, or hand-authored) against built-in or custom metrics. Writes timestamped results_<YYYYMMDD_HHMMSS>.json (consumed by eval compare) and .html (open in a browser) into the output dir, and prints a summary table to the console.

# Basic — defaults: traces from artifacts/traces/, results to artifacts/grade_results/,
# metrics from tests/eval/eval_config.yaml's metrics_to_run
agents-cli eval grade

# Advanced 1 — grade traces from a non-default location (the canonical
# pairing for `eval generate --output custom_traces/`)
agents-cli eval grade --traces custom_traces/

# Advanced 2: load metrics to run from a config file (YAML or JSON) on a specified trace file.
agents-cli eval grade --traces ./artifacts/traces/trace_1.json --config tests/eval/eval_config.yaml

See Evaluation Configuration Schema below for the config file format.

eval compare

Diffs two results_*.json files from an eval run. Run it after a fix to confirm the target metric improved without regressing others.

agents-cli eval compare baseline.json candidate.json

eval dataset synthesize

Generates user scenarios from your agent's tools and instructions, plays each against an LLM-backed user simulator, and writes graded-ready traces to artifacts/traces/ (feed straight to eval grade, skip eval generate). Invocations, flags, and compatible metrics: references/user-simulation.md.

Advanced commands

eval analyze (cluster failure modes), eval optimize (GEPA prompt tuning), and eval submit / eval results (managed cloud-side runs for CI or large datasets) are documented in references/advanced-commands.md.

Evaluation Dataset Format

An EvaluationDataset is a JSON file with an eval_cases array. Cases come in two shapes depending on how they're used:

• Inference input (what you give to eval generate) — a user prompt or a partial conversation ending in a user prompt. The agent runs and produces traces.
• Grading input (what you give to eval grade) — a complete trace including the agent's responses and tool calls. Normally produced by eval generate or eval dataset synthesize; you don't write these by hand.

See references/dataset_schema.md for the full canonical schema, all field types, and common mistakes.

Inference input format

Two shapes are supported.

(a) Simple single-turn prompt — what the scaffolded tests/eval/datasets/basic-dataset.json uses. The agent runs from scratch.

{
  "eval_cases": [
    {
      "eval_case_id": "greeting",
      "prompt": {
        "role": "user",
        "parts": [{"text": "Hello, what can you help me with?"}]
      }
    }
  ]
}

(b) Multi-turn continuation via agent_data — a partial conversation whose last turn ends with a user message; the agent's next response is evaluated. See references/dataset_schema.md (Multi-Turn / Multi-Agent Dataset) for the JSON shape.

Grading input format (traces)

A complete trace — agent responses plus function_call / function_response parts — normally produced by eval generate / eval dataset synthesize (you don't write these by hand). Authors are "user", an agent ID from the agents map, or "tool". See references/dataset_schema.md for the trace shape, multi-agent examples, and the full type reference.

Evaluation Configuration Schema

agents-cli eval run --config <path> (and eval grade --config <path>) accepts a single configuration file in either YAML (.yaml / .yml) or JSON (.json). The file declares two parts:

• metrics_to_run: the selection list of metric names to execute on this run. A name resolves to a custom_metrics entry when one matches, otherwise to the built-in metric of that name.
• custom_metrics — a definition pool of custom metrics available to this project. Defining a metric here does not run it; it must also appear in metrics_to_run (or be passed via --metrics name1,name2 on the CLI, which is equivalent to overriding metrics_to_run for that invocation).

Minimal example (YAML preferred — human-readable, no JSON escaping for prompts and Python):

metrics_to_run:
  - multi_turn_task_success     # built-in
  - example_llm_metric          # selected from custom_metrics pool below
  - agent_turn_count            # selected from custom_metrics pool below

custom_metrics:
  - name: example_llm_metric
    prompt_template: |
      Rate the agent's response 1-5 for helpfulness and accuracy.
      Prompt: {prompt}
      Final response: {response}
      Full trace (for tool-call and reasoning context): {agent_data}
      Return JSON: {"score": <1|2|3|4|5>, "explanation": "<reason>"}

  - name: agent_turn_count
    custom_function: |
      def evaluate(instance):
          turns = (instance.get("agent_data") or {}).get("turns", [])
          return {'score': len(turns)}

JSON is also accepted (same field names, with prompt_template and custom_function as escaped strings) — but always prefer YAML for human-readable configs.

Dispatch by field: custom_function → Python metric; prompt_template → LLMMetric (LLM-as-judge); neither, on a built-in name → parameterizes that built-in (e.g. metric_spec_parameters.rubric_group_key). Field reference: references/metrics-guide.md.

Agent trace field model. For datasets produced by agents-cli eval generate (or eval dataset synthesize), each eval case exposes three standard fields to a metric:

• {prompt} — the user message (or first user turn).
• {response} — the agent's final text response, extracted from the last text-bearing event. In custom_function callbacks this is instance['response'] with shape {"role": "model", "parts": [{"text": "..."}]}.
• {agent_data} — the full structured turns/events trace, useful when the judge needs to reason about tool calls or intermediate reasoning.

reference, context, and rubric_groups are yours to author on the case: eval generate carries them onto the trace but never invents them, so {reference} / {context} resolve only where you wrote them. rubric_groups is not a placeholder at all: managed rubric metrics read it off the case, and a custom_function sees instance['rubric_groups']. See references/dataset_schema.md (Per-Case Rubrics).

Code-based metrics default to local in-process execution (no GCP project or region required, but the evaluate(instance) function runs with the CLI's privileges). Set execution: "remote" on the metric to run it server-side in Vertex AI's CodeExecutionMetric sandbox instead — that path requires a configured GCP project + region.

Common Gotchas

Use Rubric-Based Tool Evaluation instead of Hardcoded Sequences

Evaluating agent tool usage using strict sequence matching is fragile because agents may call helper tools (like searches or geocoding) in different orders or perform extra proactive steps.

Instead, use multi_turn_tool_use_quality / multi_turn_trajectory_quality. These metrics automatically generate content-based and intent-based adaptive rubrics, assessing technical correctness and technical sequence logic semantically using an LLM judge rather than forcing a rigid match.

App name must match directory name

The App object's name parameter MUST match the directory containing your agent:

# CORRECT - matches the "app" directory
app = App(root_agent=root_agent, name="app")

# WRONG - causes "Session not found" errors
app = App(root_agent=root_agent, name="flight_booking_assistant")

Vertex eval region

eval run, eval grade, and eval submit default to the global endpoint. They don't inherit the manifest region (the eval services support only a subset of regions), and eval analyze is global-only. Override these per run with --region <REGION> (e.g. data residency); the service rejects an unsupported one:

400 FAILED_PRECONDITION: Unsupported region for Vertex Evaluation Service: <region>

eval generate (without the --url flag) and eval dataset synthesize run your agent locally, so they honor the agent's own .env — notably GOOGLE_CLOUD_LOCATION, which selects the model endpoint when the agent uses Vertex AI (GOOGLE_GENAI_USE_VERTEXAI=true); it's unused with a GEMINI_API_KEY (AI Studio). They take no --region and never override your .env with the manifest region; change the model region by editing .env. One caveat for synthesize: its scenario-generation step is a server-side eval call at GOOGLE_CLOUD_LOCATION, so keep that an eval-supported region (global by default) even though the agent itself could run elsewhere.

No eval region fits your data-residency rules? Fall back to local custom metrics — a custom_metrics entry with a custom_function (execution: local, the default) grades in-process with no GCP region required. You lose the managed built-in metrics, but your custom_function can still call an LLM judge in a compliant region itself — so LLM-as-judge grading stays available anywhere.

The before_agent_callback Pattern (State Initialization)

Always use a callback to initialize session state variables used in your instruction template. This prevents KeyError crashes on the first turn:

async def initialize_state(callback_context: CallbackContext) -> None:
    state = callback_context.state
    if "user_preferences" not in state:
        state["user_preferences"] = {}

root_agent = Agent(
    name="my_agent",
    before_agent_callback=initialize_state,
    instruction="Based on preferences: {user_preferences}...",
)

Model thinking mode may bypass tools

Models with "thinking" enabled may skip tool calls. Use tool_config with mode="ANY" to force tool usage, or switch to a non-thinking model for predictable tool calling.

Common Eval Failure Causes

Proving Your Work

Don't assert that eval passes — show the evidence. Concrete output prevents false confidence and catches issues early.

• After running eval: Paste the scores table output so the user can see exactly what passed and failed.
• After fixing a failure: Show before/after scores for the specific case you fixed, and confirm no other cases regressed.
• Before deploy: Rerun agents-cli eval run and show the scores for every case, not just the one you fixed. eval run exits 0 whatever the scores are, so the numbers you paste are the gate, not the exit code.

Related Skills

• /google-agents-cli-workflow — Development workflow and the spec-driven build-evaluate-deploy lifecycle
• /google-agents-cli-adk-code — ADK Python API quick reference for writing agent code
• /google-agents-cli-scaffold — Project creation and enhancement with agents-cli scaffold create / scaffold enhance
• /google-agents-cli-deploy — Deployment targets, CI/CD pipelines, and production workflows
• /google-agents-cli-observability — Cloud Trace, logging, and monitoring for debugging agent behavior