---
name: agent-testing
description: Validate non-deterministic AI agents - chatbots, voice agents, phone/calling agents, and image generators - with TestMu AI's Agent Testing Platform. AI evaluators hold real conversations with your agent, score chat/voice quality metrics and 30+ call metrics, and return a Green/Yellow/Red go-live verdict. Covers agent connection contracts, the web workflow, and CI/CD automation.
---

# Agent Testing Platform - AI-Agent Evaluation Skill

Use the **Agent Testing Platform** to test **AI agents** - chatbots, voice agents, phone/calling agents, and image generators - the way a real user would. AI agents are **non-deterministic**: ask the same question twice and you get two different answers, so there is **no selector to assert against and no DOM state to check**. Instead, an **AI evaluator holds a full multi-turn conversation with your agent and scores every response**, rolling up into a **production-readiness verdict** before you ship.

> "We use AI agents to test other AI agents."

**Two ways to drive it:**
1. **Web platform** - create an agent, upload context, generate scenarios, run the evaluation, read results (Section 5).
2. **Automation** - a REST API with SSE status streaming, an in-platform cron scheduler, and the `testmu-a2a-cli` for terminal/CI use (Section 11).

**When to use this vs. KaneAI / kane-cli:** use **Agent Testing** to validate a **model/agent's conversational behavior** (does it hallucinate? is it biased? does it stay on topic?). Use **KaneAI** or **`kane-cli`** to test a **deterministic web/mobile UI**. Selenium asserts deterministic state; this platform replaces that with reproducible, evidence-backed scoring across generated scenarios.

---

## 1. Decision Tree

**What are you validating?**
- A **text chatbot** (HTTP endpoint) -> **Chat** agent (Sections 4-8)
- A **voice assistant** (audio over an endpoint - no phone) -> **Voice** agent (Section 9a)
- An **inbound phone / IVR agent** (real calls to its number) -> **Phone Caller Inbound** (Section 9b)
- An **outbound dialer / outreach agent** -> **Phone Caller Outbound** (Section 9c)
- **AI-generated images** against prompts/specs/brand rules -> **Image Analyzer** (Section 10)
- **Production call recordings** (batch analysis) -> Post-evaluation (Section 9d)
- **Security / adversarial robustness** -> Red-teaming (Section 12)
- **CI/CD deployment gate** -> REST API / SSE / CLI (Section 11)

**Every chat/voice/phone evaluation follows the same shape:**
Connect the agent -> generate scenarios from your docs -> AI evaluator converses with the agent -> responses scored against metrics -> **Go-Live verdict** (Green / Yellow / Red).

---

## 2. Core Concepts

| Term | Meaning |
|---|---|
| **AI evaluator / simulator** | Converses with your agent-under-test like a real user - following the scenario goal, adapting to responses, including interruptions, clarifications, topic pivots, and edge-case inputs. |
| **Scenario** | A generated test case with a persona, an objective, and behavioral guardrails. Each is assigned a **complexity and risk level**, organized into behavioral **test categories** (Section 6). |
| **Metric** | A scored quality dimension. Chat/voice metrics are scored **0-100%**; pass/fail derives from **configurable per-metric thresholds** (0.0-1.0, with per-metric directionality). |
| **Validation criteria** | Custom, evidence-based pass/fail rules; results are **Pass / Fail / Unable to Verify** per criterion, each with **High/Medium/Low confidence** and supporting evidence. |
| **Go-Live verdict** | **GREEN** (score >= 80) Ready for Production, **YELLOW** (65-79) Ready with Caveats, **RED** (< 65) Not Ready, **NO_DATA** insufficient data. Derived from a weighted composite **Overall Score (0-100)** (Section 8). |
| **Confidence level** | Attached to the Go-Live assessment, based on evaluations run: **HIGH** 100+, **MEDIUM** 50-99, **LOW** 20-49, **VERY_LOW** < 20. |

**Timing expectations** (for orchestration/polling): project setup 5-10 min, document upload ~5 min, scenario generation (**60-100+ scenarios**, streamed live to the UI) 3-8 min, first test run 20-40 min (suite-dependent), threshold configuration ~15 min.

**Cost tracking:** per-test execution costs (call minutes, AI evaluation compute) are tracked and aggregated at project, suite, and organization level.

---

## 3. Agent Types (5)

| Agent type | What it tests | How it connects | Headline evaluation |
|---|---|---|---|
| **Chat** | Text chatbots | HTTPS endpoint (HTTP POST) | Multi-turn conversation, **9 quality metrics** |
| **Voice** | Voice assistants | Same endpoint mechanics as Chat, but conversations happen as **audio (WAV)** instead of text | Same 9 metrics, scored on the audio transcript |
| **Phone Caller Inbound** | IVR / inbound support bots | **Real phone calls** - the platform dials your agent's number | **30+ call metrics** in 8 categories |
| **Phone Caller Outbound** | Outbound dialers / outreach agents | The platform **provisions a recipient number**; your agent is triggered to call it | Same call metrics + outbound-specific features |
| **Image Analyzer** | AI-generated images | Image upload / URL + the original prompt | **Quality Score (0-100)** + criteria compliance |

> **Voice != Phone.** A Voice agent is functionally identical to Chat - endpoint-based, WAV audio instead of text, same 9 metrics. Only the two Phone Caller types place real phone calls and use the 30+ call metrics.

**Feature availability boundaries** (don't attempt unsupported operations):
- Workflow & document upload, **Endpoint Profiles**, **Playground**, HyperExecute integration, profile import/export -> **Chat & Voice only**.
- AI scenario generation -> Chat/Voice (doc-driven), Inbound (**up to 20**), Outbound (**up to 7**); **not** Image Analyzer.
- Phone numbers, voice selection, noise simulation, live calls, recording upload/playback, DTMF -> **Phone Caller only**; passive mode & number pool -> **Outbound only**.
- **Image Analyzer** has no test suites, personas, metric thresholds, go-live assessment, validation criteria, or scheduled runs.

---

## 4. Connecting Your Agent

### What you provide, per agent type

| Agent type | Required |
|---|---|
| **Chat / Voice** | HTTPS **endpoint URL** + any required **auth headers/credentials**. Multi-step auth flows (login -> session token -> API call) are supported via **Endpoint Profiles**. A **bot/assistant identifier** if one API hosts multiple bots. |
| **Phone Caller** | The **E.164-formatted phone number** your agent answers on (inbound) or dials from (outbound). No changes to your telephony stack required. |
| **Image Analyzer** | The images (upload or URL) and the **original generation prompt**; optionally evaluation criteria. |

### The HTTP request contract (Chat/Voice)

The platform calls your chatbot with standard **HTTP POST** requests:

```bash
curl -X POST https://api.examplechatbot.com/chat \
  -H "Authorization: Bearer sk-example-a1b2c3d4e5f6" \
  -H "Content-Type: application/json" \
  -d '{ "assistantId": "asst_7xG9kPqR2mN4", "input": "Hi, I need help with my account" }'
```

- Field names (`assistantId`, `input`) are illustrative - your API may use `message`, `query`, `botId`, etc. **The platform adapts to whatever request structure your API expects**, and can match a custom body format exactly.
- **Additional headers are fully supported, unlimited** (session ID, API version, workspace ID, tracking headers) - every configured header is forwarded with every request.

**Response contract** - the platform parses your schema; the key field:

```json
{
  "id": "1310ab59-...",
  "input":  [{ "role": "user",      "content": "Hi, I need help with my account" }],
  "output": [{ "role": "assistant", "content": "Sure! Could you please provide more details..." }],
  "createdAt": "2025-10-29T06:02:45.616Z",
  "cost": 0.0063
}
```

- **`output` / `response` is the primary field the platform evaluates** - nested under `role: "assistant"` or a top-level string.
- **Multi-turn:** the platform reads `output` from each response and continues the dialogue across exchanges. The conversation loop: scenario selected -> HTTP request -> agent responds -> platform continues the turn based on the scenario goal and prior context -> an automated end condition (**goal met, max turns reached, or unrecoverable error**) terminates the session -> metrics scored.

### Connection methods (network reachability)

| | **A: Public API** | **B: Secure Proxy** | **C: Localhost** |
|---|---|---|---|
| Agent reachable from internet? | Yes | No (VPC / firewall) | No (dev machine) |
| Proxy agent needed? | No | Yes - installed in your network | Yes - installed on your machine |
| Firewall changes | None | None (**outbound-only tunnel**) | None |
| Best for | Production / cloud bots | Enterprise / on-premise | Dev & staging (`http://localhost:3000/chat`) |

The lightweight **proxy agent** (provided by the platform) handles *network reachability only* - it never bypasses your chatbot's own authentication. **Credentials are encrypted at rest, decrypted only at runtime, and never exposed in reports or logs.**

### Endpoint Profiles (Chat/Voice)

The full connection configuration lives in an **Endpoint Profile**:
- **Postman collection import** (with optional environment files; supports nested folders and variable substitution) or **manual JSON** (URL, method, headers, body).
- **Multi-phase execution**: `suite_setup` (login/auth) -> `scenario_setup` (session creation) -> `chat` (conversation).
- **Variable management**: static values, auto-generated values (UUID, mobile number, timestamp, email), and values extracted from API responses.
- **Retry & caching** (cache suite-setup results), **Test Endpoint** dry-run to verify connectivity before burning an evaluation, JSON import/export across projects, default-profile marking.

### Playground (pre-flight, Chat/Voice)

Before running evaluations, verify the wiring interactively: send real messages to your configured endpoint, hold multi-turn conversations, test connectivity via **cURL verification**, and let the platform **auto-detect your request/response schema**.

---

## 5. Web Workflow - Test Your First AI Agent

1. Click **Create Agent** in the left sidebar; give the agent a **name** and **description**, then click **Create Agent**.
2. **Upload the required documents** so the platform understands your agent's requirements.
3. **Select test categories** to generate (Section 6) - e.g. **Personality & Tone** checks that the agent responds professionally. Select multiple categories to widen coverage.
4. Click **Generate Test Scenarios** - a team of specialised AI agents generates scenarios **in parallel** (streamed live, typically 3-8 minutes).
5. Review the scenarios: **filter by category**; each carries a **complexity and risk level**.
6. Enter your **bot/agent's API URL** (or select the Endpoint Profile, Section 4) and click **Run Evaluation**.
7. Open the **Evaluation Results** tab for the breakdown across metrics - e.g. whether the conversation was relevant and whether the agent stayed on topic - plus transcripts and the Go-Live assessment (Section 8).

> **Voice bots:** the platform tests them by automatically generating **audio** for the test cases and evaluating the **transcribed** audio responses.

---

## 6. Scenario Generation

- Scenarios are **auto-generated from your uploaded context** - PRDs, docs, knowledge bases - typically **60-100+ scenarios** for chat/voice agents, spanning happy paths, edge cases, and adversarial inputs. Phone agents have hard generation limits: **up to 20 (inbound)** / **up to 7 (outbound)** per generation, with configurable personas, languages, and special instructions.
- Generated scenarios are organized across **16 behavioral test categories**: Conversational Flow, Intent Recognition, Context & Memory, Multi-Turn Reasoning, User Experience, Personality & Tone, Proactive Behavior, Multimodal Interactions, Third-Party Integration, Error Handling, Consistency, Security, Compliance & Governance, Data Privacy, Performance, Recovery Mechanisms.
- Every scenario includes a **caller/user persona, a conversation objective, and expected behavioral guardrails**, plus a complexity and risk level.
- **Test Profiles & Personas** - inject reusable key-value test data with typed fields (**string, number, boolean, email, URL, textarea, JSON**) and use the pre-built or custom **persona library** for targeted scenario execution.
- **Validation Criteria** - define custom, evidence-based pass/fail rules; each result reports **Pass / Fail / Unable to Verify** with evidence and a High/Medium/Low confidence level.

---

## 7. Quality Metrics Reference

The evaluation engine scores every completed conversation against a per-agent-type metric contract:

### Chat & Voice - the 9 quality metrics (each scored 0-100%)

1. **Hallucination Detection** - fabricated or incorrect information
2. **Bias Detection** - biased or unfair responses
3. **Completeness** - fully addressing the user's request
4. **Context Awareness** - using conversation context correctly
5. **Response Quality** - overall quality of individual responses
6. **Conversation Flow** - natural, coherent dialogue progression
7. **Tone Consistency** - maintaining a consistent tone and persona across the conversation
8. **Positive User Outcome** - whether the user ends the conversation with their goal achieved
9. **Root-Cause Understanding** - addressing the user's underlying problem, not just surface symptoms

Every metric carries a **High / Medium / Low confidence level** based on evaluation volume. Failing transcripts are annotated with the evidence that drove each score.

**Per-metric thresholds:** set a minimum acceptable score (0.0-1.0) per metric, configure whether higher or lower is better, save **named configurations** (e.g. "Strict", "Default"), and toggle metrics active/inactive.

**Every evaluation result includes:** the Overall Score (weighted average), per-metric score with a **pass/fail badge** against your thresholds, per-metric analysis text, the full multi-turn transcript, identified strengths, areas for improvement, actionable recommendations, and validation-criteria results.

### Phone Callers - 30+ call metrics in 8 categories

| Category | Metrics |
|---|---|
| **A. Conversation Flow & Interaction Dynamics** | Average Latency (ms - response time after user stops speaking), Words Per Minute, AI Talk Ratio (%), User Talk Ratio (%), AI Interrupting User (%), User Interrupting AI (%) |
| **B. Accuracy & Effectiveness** | First Call Resolution / FCR (%), Intent Recognition Accuracy (%), Task Completion Success Rate (%), Instruction Following (%), Response Consistency (%) |
| **C. User Experience & Satisfaction** | CSAT (%), CSAT Reason (text), User Sentiment (text), Early Termination (%) |
| **D. Business Operational** | Containment Rate (% resolved without human escalation), AI-to-Human Handoff Rate (%) |
| **E. Audio Voice Quality** | Average Pitch (Hz; normal 85-300), Voice Quality Index (0-5 composite), Signal-to-Noise Ratio (%) |
| **F. Speech-to-Text Evaluation** | STT Accuracy (%), STT Verdict (Pass/Fail), STT Summary, Mismatch Examples |
| **G. Validation Results** | Compliance rate against custom criteria, Pass/Fail/Unable-to-Verify counts per criterion |
| **H. Detected Issue Tags** | see below |

**Threshold bands (selected):** Average Latency - Excellent <= 1000 ms, Good <= 2500 ms, Poor > 2500 ms, Words Per Minute - >= 160 fast, 131-160 good, < 110 slow, Voice Quality Index - >= 2.5/5 pass, Average Pitch - 85-300 Hz normal.

**Detected issue tags** - auto-flagged in every call recording: Latency Issues, Hallucination in Call Flow, Transcript Issue, Patchy Audio, Running in Loop, Incorrect STT, Interruption Handling, Number Issue, Background Noise, No Response, BLANK/EMPTY STT.

**Metric configuration:** enable/disable whole metric categories or individual metrics per project - skipping non-critical metrics reduces analysis time and cost.

> The 9 chat/voice metrics apply to **Chat and Voice only**; the call-metric categories apply to **Phone Callers only**. Validation Criteria apply to Chat, Voice, and Phone Callers.

---

## 8. Reading Results & the Go-Live Verdict

The **Go-Live Assessment** aggregates all evaluated scenarios into a production-readiness verdict (available for Chat, Voice, and Phone Callers - **not** Image Analyzer):

| Verdict | Score | Meaning |
|---|---|---|
| **GREEN** | >= 80 | **Ready for Production** - critical thresholds met; consistent, reliable behavior |
| **YELLOW** | 65-79 | **Ready with Caveats** - most thresholds met; secondary metrics below target; conditional approval with identified risk areas to monitor |
| **RED** | < 65 | **Not Ready** - one or more critical thresholds unmet; failure categories surfaced for remediation |
| **NO_DATA** | - | Insufficient data to assess |

**The assessment contains:**
- **Overall Score** - weighted composite (0-100).
- **Confidence level** - based on evaluations run: HIGH 100+, MEDIUM 50-99, LOW 20-49, VERY_LOW < 20.
- **Dimension Scores** - Functional Completeness, Quality Standards, Risk Profile, Operational Readiness (each weighted 25%).
- **Scenario Coverage** - well-tested vs. untested (vs. high-risk for phone) scenarios.
- **Risk Assessment** - AI-powered failure-pattern analysis with prioritized action items.
- **Validation Criteria Summary** - aggregated compliance rate.
- **AI Insights** - actionable recommendations.

Thresholds that drive the verdict are **configurable per project and per metric** (e.g. a regulated deployment might require Compliance >= 0.95 and Data Privacy >= 0.98); threshold profiles are **version-controlled and can differ per environment** (development, staging, production). Track score changes across runs to catch **model drift**.

---

## 9. Voice & Phone Agent Testing

### 9a. Voice agents (no phone involved)

Functionally identical to Chat with one difference: **conversations happen as audio (WAV) instead of text**. The agent's voice responses are captured, transcribed, and evaluated on the full conversation transcript against the same 9 metrics. Voice testing supports **REST, WebSocket & WebRTC** transports.

### 9b. Phone Caller Inbound

- **Direction:** the platform's simulated caller **dials your agent's number** and follows the scenario script - adapting dynamically, including interruptions, clarifications, topic pivots, and edge cases.
- **Per-scenario voice configuration:** voice selection from a library with audio preview (multiple providers and accents), **background sound** (15 presets: cafe, street, factory, rain, crowd, market, train, radio interference, ...), **response timing** (0.5-5.0 s wait after speech ends - handles agents that speak in multiple sentences), **max call duration** (60-1800 s), **first speaker** (simulated user or agent; inbound default: simulator).
- **Phone number management:** register your agent's numbers with country-code selection (**20+ countries**), set a default number, numbers displayed masked.
- **Agent Profiles:** reusable org-level caller personas (name, phone number, voice, background noise) with active/inactive toggle.
- **Live monitoring:** call status, live duration counter, and the ability to terminate a call in progress.
- **Call lifecycle:** full audio captured and stored securely; the call ends on scenario completion or the configured max duration; the evaluation engine transcribes and scores the recording **within minutes** of completion.
- **Telephony-agnostic:** configure your own voice-provider credentials in settings (encrypted at rest); existing contracts and numbers are reused - no platform-level telephony commitment.

### 9c. Phone Caller Outbound

Everything from Inbound, plus these outbound-specific differences:
- **Direction reversed:** the platform **reserves a recipient number from its outbound pool**, and your agent is triggered to call it - the platform answers and evaluates.
- **Scenario generation capped at 7** (vs. 20 inbound); select an outbound **caller profile** during generation.
- **First speaker default:** the **agent** speaks first (vs. simulator for inbound).
- **Passive Mode:** listen to live outbound calls **without interfering** - QA monitoring.
- **Pool management:** view available numbers, manage per-suite reservations, clear reservations when done.

### 9d. Post-evaluation: production recording analysis (Phone Callers)

Two evaluation modes exist: **pre-evaluation** (live simulated test calls, above) and **post-evaluation** - upload real production recordings:
- **Formats:** MP3, WAV. Upload transcripts alongside recordings for **STT comparison**. **Batch upload** analyzes multiple recordings in parallel; choose metric categories selectively.
- Organize with bookmarking, tagging, and search/filter (name, status, date, tags).
- **Recording playback:** audio player with duration tracking, **speaker-identified transcript** (agent vs. user), **DTMF detection** - keypad inputs (0-9, *, #) captured and displayed in the transcript, downloadable audio + transcript.

---

## 10. Image Analyzer Agent

Score AI-generated images against the prompt that produced them:

- **Inputs:** upload an image or provide a URL **plus the original prompt**. Formats: **JPG, JPEG, PNG, GIF, WEBP, BMP**, max **20 MB** per image. **Batch analysis: up to 50 images** with a shared prompt. Analysis status: Pending / Completed / Failed.
- **Results per image:** **Quality Score (0-100)** for overall quality and prompt adherence, **Matches** (elements that correctly match the prompt), **Discrepancies** (missing or incorrect elements), Overall Assessment, Detailed Observations.
- **Score interpretation:** 90-100 Excellent (Green), 80-89 Good (Green), 60-79 Fair (Yellow), 0-59 Poor (Red).
- **Custom Evaluation Criteria** (three types): **Brand Guidelines** (allowed/prohibited colors, required fonts, logo requirements), **Technical Specifications** (dimensions, aspect ratio, formats, max file size, min resolution), **Custom Rules** (freeform text, checklist items). Each active criterion reports **PASS / FAIL / PARTIAL** with details.
- **Analytics dashboard:** average/highest/lowest scores, total analyses, daily score breakdown over the **last 30 days**, and the **top 20 prompts** ranked by average quality score.

> No test suites, personas, thresholds, go-live verdicts, or scheduled runs for this agent type - it's a scoring surface, not a conversation evaluator.

---

## 11. Automation: API, Scheduler & CLI

### REST API + SSE (the documented platform path)

- The platform exposes a **REST API for all core operations** (Bearer / API-key auth) - trigger suite executions from any CI/CD pipeline and receive structured pass/fail signals.
- **Real-time status streams via server-sent events (SSE)** for live build dashboards.
- An in-platform **cron scheduler** runs autonomous scheduled regressions independent of deployments.

### testmu-a2a-cli (terminal / CI)

```bash
pip install testmu-a2a-cli          # Python 3.10+
testmu-a2a auth --username YOUR_USERNAME --access-key YOUR_KEY
```

**Evaluate a chatbot:**
```bash
testmu-a2a test \
  --agent https://your-chatbot-endpoint.com \
  --spec "E-commerce customer support chatbot" \
  --count 200 \
  --format json \
  --output results.json
```

**Red-team an agent:**
```bash
testmu-a2a redteam --agent https://your-chatbot-endpoint.com --output redteam-results.json
```

**Test a phone agent:**
```bash
testmu-a2a call --agent https://your-phone-agent-endpoint.com --type inbound --output call-results.json
```

**Config-driven (CI/CD):**
```bash
testmu-a2a init                          # generates testmu-a2a.yaml
testmu-a2a run --config testmu-a2a.yaml  # commit the config, run in CI
```

Results emit as **structured JSON** (and JUnit XML for CI platforms); gate the pipeline on the exit code. The CLI grades the same nine-metric chat contract as Section 7.

---

## 12. Red-Teaming / Security

The `redteam` surface probes your agent for adversarial and safety failures before attackers do:

- Prompt injection
- Jailbreak attempts
- Data exfiltration
- PII leakage
- (...and more)

The **Security** and **Compliance & Governance** behavioral categories (Section 6) generate adversarial scenarios in standard evaluations too, and validation criteria let you encode compliance rules with evidence-based scoring. Findings roll into the same scores and verdict (Section 8).

---

## 13. Context Inputs & Connectors

**Multi-modal document processing** - what you can upload as agent context (processed with multiple LLMs):

| Modality | Handling |
|---|---|
| Text | PDF, DOCX, XLSX, TXT, MD |
| Images | with **OCR** |
| Audio | with **speech-to-text** |
| Video | with **frame extraction and transcription** |

- **Structured requirement extraction** from unstructured data: entities, business rules, integration points, and security requirements - with source-requirement ID traceability and cross-document linkage.
- **Multilingual requirement analysis** - extract and analyze requirements from documents in multiple languages; supports testing agents that operate across languages.

**Out-of-the-box data source connectors (3)** - private/self-hosted instances supported via **secure tunnels**:

| Connector | Pulls |
|---|---|
| **GitHub** | Repository content, README retrieval, file access (public & private repos) |
| **JIRA** | Ticket retrieval, project info, issue tracking (incl. private instances) |
| **Confluence** | Page content, documentation, knowledge bases (incl. private instances) |

**Platform capabilities:** project & environment management (bulk creation, scoped variables), **model governance** - controlled workflows, version tracking, access controls, audit trails, policy-aligned oversight across GenAI model lifecycles, scheduling engine, unified dashboards, exportable reports, and real-time pass/fail trends.

---

## 14. FAQs & Reference

**What is AI agent testing?** Validating non-deterministic AI agents (chatbots, voice assistants, phone agents) against standardized quality metrics like hallucination, bias, completeness, and context awareness.

**How does it differ from Selenium / manual QA?** Selenium asserts deterministic state (button text equals "Submit"). AI agents are non-deterministic - no selector to check, and manual review of conversation logs doesn't scale. This platform replaces both with reproducible, evidence-backed scoring.

**Agent testing vs. agentic testing?** *Agent testing* validates AI agents **you build**. *Agentic testing* is QA **performed by an AI agent** that plans, authors, and runs your software tests (that's KaneAI).

**Do I need to change my chatbot's code?** No - the platform adapts to your request/response schema; you provide the endpoint, credentials, and headers (Section 4). Firewalled or localhost agents use the outbound-only proxy agent.

**Key numbers:**

| Item | Value |
|---|---|
| Agent types | 5 (Chat, Voice, Phone Inbound, Phone Outbound, Image Analyzer) |
| Chat/voice quality metrics | 9 (scored 0-100%) |
| Phone call metrics | 30+ across 8 categories |
| Behavioral test categories | 16 |
| Auto-generated scenarios (chat/voice) | 60-100+ (3-8 min, streamed) |
| Phone scenario generation | up to 20 inbound / 7 outbound |
| Go-Live verdict | GREEN >= 80, YELLOW 65-79, RED < 65, NO_DATA |
| Assessment confidence | HIGH 100+, MEDIUM 50-99, LOW 20-49, VERY_LOW < 20 evaluations |
| Background-noise presets | 15 |
| Response timing / max call duration | 0.5-5.0 s / 60-1800 s |
| Image batch | up to 50 images, 20 MB each |
| First test run | 20-40 minutes |

**Links:**
- **Docs:** https://www.testmuai.com/support/docs/getting-started-with-agent-testing-platform/, https://www.testmuai.com/support/docs/testing-your-first-ai-agent/, https://www.testmuai.com/support/docs/chat-agent-api-integration/
- **Product:** https://www.testmuai.com/agent-testing/, https://www.testmuai.com/agent-to-agent-testing/
- **Community:** https://community.testmuai.com/, **Support:** support@testmuai.com