AgentEval

Open-source multi-turn AI agent simulation and evaluation.

Define a test scenario in YAML, run it against your agent's HTTP endpoint, and get a structured pass/fail report — from your terminal, for free, with no mandatory cloud dependency.

pip install agent-eval-cli
agenteval run test_cases/ --mode scripted

---

The gap this fills

Every existing tool evaluates what your agent already said. You bring the conversation; they score it.

Nobody simulates the conversation for you — free — against any endpoint.

Tool	Multi-turn simulation	Free	Open source
LangSmith	❌	❌	❌
Braintrust	❌	❌	❌
Promptfoo	❌	✅	✅
DeepEval	❌	✅	✅
Coval	✅	❌	❌
AgentEval	✅	✅	✅

---

Install

# Core — scripted mode only, no ML scorers
pip install agent-eval-cli

# + Groq simulation mode (free API key, no credit card)
pip install agent-eval-cli[groq]

# + ML scorers (sentence-transformers, ~500MB)
pip install agent-eval-cli[ml]

# Everything
pip install agent-eval-cli[all]

Requires Python 3.9+

---

Quick start

1. Write a test case:

# test_cases/booking_happy_path.yaml
meta:
  scenario_id: "booking-happy-path-001"
  name: "Patient successfully books a doctor's appointment"
  tags: ["healthcare", "booking"]

agent:
  endpoint: "http://localhost:8080/chat"
  method: "POST"
  headers:
    Content-Type: "application/json"
  request_template: |
    {
      "session_id": "${SESSION_ID}",
      "message": "${USER_MESSAGE}"
    }
  response_path: "response"
  timeout_seconds: 30

user_persona:
  name: "Sarah Chen"
  tone: "polite but slightly anxious"
  background: "First-time patient, not tech-savvy"
  opening_message: "Hi, I need to see a doctor about my knee."

outcome_type: success   # success | refusal | escalation

conversation:
  max_turns: 12
  min_turns: 3
  expected_turns: 8

evaluation:
  success_intent: "User receives a confirmed appointment with specific date, time, and doctor name"
  success_keywords:
    - "appointment confirmed"
    - "booked for"
    - "confirmation number"
  forbidden_phrases:
    - "I don't know"
    - "I cannot help"
  context_facts:
    - "Appointments available Monday through Friday"
    - "Dr. Smith specialises in orthopedics"
    - "Clinic hours are 8am to 6pm"
  thresholds:
    task_completion:       0.65
    instruction_following: 0.80
    turn_efficiency:       0.60
    aggregate:             0.70

scripted_turns:
  - "Hi, I need to see a doctor about my knee. It's been hurting for two weeks."
  - "Is Dr. Smith available this week?"
  - "Thursday at 2pm works for me."
  - "My name is Sarah Chen, date of birth March 12 1990."
  - "Yes, that email is correct."
  - "Great, thank you so much."

2. Run:

agenteval run test_cases/ --mode scripted

3. View results:

agenteval dashboard ./reports/

---

Simulation modes

--mode scripted — deterministic CI mode

Sends the fixed scripted_turns from your YAML in order. Zero API calls. Zero variance. After the last scripted turn, automatically signals goal completion if the agent responded without error.

Use this for CI/CD gates — results are reproducible across every run.

agenteval run test_cases/ --mode scripted --concurrency 4 --fail-on-threshold 0.70

--mode groq — realistic simulation

Llama 3.3 70B (via Groq's free API) plays the user side, guided by the persona and outcome type you defined. Has natural variance — useful for exploratory evaluation and finding edge cases scripted mode misses.

Requires a free Groq API key — no credit card.

export GROQ_API_KEY=your_key_here
agenteval run test_cases/ --mode groq

Privacy notice: when using --mode groq, your context_facts are sent to Groq's API. Use --mode scripted for sensitive or proprietary data.

---

Scorers

All scoring runs locally after the conversation completes. No API calls.

These are heuristic signals designed to catch obvious failures — not authoritative ground truth. Calibrate thresholds against real transcripts before using them as hard CI gates.

Scorer	What it measures	Requires [ml]
Task Completion	Semantic similarity between agent responses and the outcome intent	Yes
Instruction Following	Did the agent obey forbidden/required phrase rules?	No
Response Coherence	Are agent responses contextually relevant to what was asked?	Yes
Turn Efficiency	Did the agent resolve the goal in a reasonable number of turns?	No
Hallucination Risk	Did the agent state facts not grounded in context_facts?	Yes

Scorers 1, 3, and 5 return None when [ml] is not installed — they are excluded from the aggregate and weights are re-normalised automatically.

Default aggregate weights:

Scorer	Weight
Task Completion	30%
Instruction Following	25%
Response Coherence	20%
Turn Efficiency	15%
Hallucination Risk	10%

---

CLI reference

# Run test cases (scripted mode — CI-safe, no API key)
agenteval run test_cases/ --mode scripted

# Run with Groq simulation
agenteval run test_cases/ --mode groq

# Filter scenarios by tag
agenteval run test_cases/ --mode scripted --tag customer-support

# Parallel execution (default: 4)
agenteval run test_cases/ --mode scripted --concurrency 8

# Fail with non-zero exit code when aggregate score is below threshold
agenteval run test_cases/ --mode scripted --fail-on-threshold 0.70

# Save reports to a directory (auto-timestamped filenames)
agenteval run test_cases/ --mode scripted --output-dir ./reports/

# Print full conversation transcript on failure
agenteval run test_cases/ --mode scripted --verbose

# Validate YAML test cases without running
agenteval validate test_cases/booking.yaml
agenteval validate test_cases/

# Open dashboard for all reports in a directory
agenteval dashboard ./reports/

# Version
agenteval --version

---

YAML test case schema

outcome_type

Three outcome types are supported:

• success — the agent should complete the user's request
• refusal — the agent should correctly decline (e.g. no availability, policy block)
• escalation — the agent should route the user to a human

Refusal example:

outcome_type: refusal
policy_reason: "No Sunday appointments available per clinic policy"

evaluation:
  refusal_intent: "Agent clearly communicates no availability without offering to book an unavailable slot"
  refusal_keywords:
    - "unfortunately"
    - "no availability"
    - "fully booked"
  forbidden_phrases:
    - "appointment confirmed"
    - "I can book that"

scripted_turns:
  - "I need to book an appointment for next Sunday."
  - "What about Monday at 9pm?"
  - "Is there anything at all this week?"

Environment variable substitution

The following variables are substituted into request_template at runtime:

Variable	Value
${SESSION_ID}	UUID4 generated per scenario — prevents state leakage on stateful agents
${USER_MESSAGE}	The current user turn content
${AGENT_API_KEY}	Resolved from environment (set in .env or CI secrets)

Any ${VAR} pattern in the template is resolved from the environment.

response_path

Uses jmespath dot notation to extract the agent's reply from the JSON response body.

response_path: "response.text"       # {"response": {"text": "..."}}
response_path: "choices[0].message.content"  # OpenAI-style
response_path: "reply"               # {"reply": "..."}

---

CI/CD — GitHub Action

# .github/workflows/agenteval.yml
name: AgentEval — AI Agent Tests

on:
  pull_request:
    branches: [main]
  push:
    branches: [main]

jobs:
  agent-eval:
    runs-on: ubuntu-latest
    timeout-minutes: 15

    steps:
      - uses: actions/checkout@v4

      - uses: actions/setup-python@v5
        with:
          python-version: "3.11"

      - name: Cache pip
        uses: actions/cache@v4
        with:
          path: ~/.cache/pip
          key: ${{ runner.os }}-pip-${{ hashFiles('**/pyproject.toml') }}

      - name: Install AgentEval
        run: pip install "agent-eval-cli==0.1.0"

      - name: Start agent under test
        run: docker compose up -d agent   # replace 'agent' with your service name

      - name: Wait for agent readiness
        run: |
          for i in {1..12}; do
            if curl -sf http://localhost:8080/health > /dev/null 2>&1; then
              echo "Agent ready."; exit 0
            fi
            echo "Attempt $i/12 — retrying in 5s..."; sleep 5
          done
          echo "Agent did not become ready in 60s."; exit 1

      - name: Run AgentEval
        env:
          AGENT_API_KEY: ${{ secrets.AGENT_API_KEY }}
        run: |
          agenteval run test_cases/ \
            --mode scripted \
            --concurrency 4 \
            --output-dir ./reports/ \
            --fail-on-threshold 0.70

      - name: Upload reports
        uses: actions/upload-artifact@v4
        if: always()
        with:
          name: agenteval-reports-${{ github.sha }}
          path: ./reports/
          retention-days: 30

Reports are uploaded as a build artifact on every run. Use actions/download-artifact in a subsequent step to post results as a PR comment.

---

Demo agent (try it in 2 minutes)

A ready-made customer support agent for TableEase (a restaurant booking service) is included so you can run AgentEval end-to-end without building your own agent first.

# Install dependencies
cd demo_agent
pip install -r requirements.txt

# Set your Groq API key (free at console.groq.com)
export GROQ_API_KEY=your_key_here

# Start the agent
uvicorn main:app --reload
# Agent running at http://localhost:8000/chat

Then in a second terminal, run the included test cases:

agenteval run test_cases/ --mode scripted --output-dir ./reports/
agenteval dashboard ./reports/

The demo agent handles table bookings, cancellation policy questions, competitor refusals, and escalations to a human manager — covering all three outcome_type values.

---

Dashboard

agenteval dashboard ./reports/
# Opens http://localhost:8080 in your browser

Features:

• Run summary with pass/fail counts and aggregate score
• Radar chart showing all 5 scorer dimensions per scenario
• Per-scenario score breakdown with pass/fail per scorer
• Full conversation transcript replay (chat bubble view)
• Evidence items flagged by instruction-following and hallucination scorers
• Multi-report dropdown to switch between runs

---

v1 scope

v1 supports synchronous JSON over HTTP POST only.

Out of scope for v1 (planned for v2):

• WebSocket / SSE / streaming responses
• OAuth and cookie-based authentication
• Multi-message response payloads
• Non-JSON payloads

---

License

MIT