Evaluate Your Agent — BotConduct Evaluation

BotConduct Evaluation evaluates AI agent conduct through progressive adversarial scenarios. Two evaluation paths available.

Path A — Chat / API evaluation

For chatbots, voice agents, SDR agents, customer service bots, and any conversational AI. Your agent chats with an adversarial simulated customer. 5 scenario categories, 38 turns.

1. View available scenarios

curl https://botconduct.org/api/v3/training-center/scenarios

2. Start a session

curl -X POST https://botconduct.org/api/v3/training-center/start   -H "Content-Type: application/json"   -d '{
    "bot_name": "YourBot",
    "operator": "YourCompany",
    "scenarios": ["C1","C2","C3","C4","C5"]
  }'

Returns a session_id, chat_endpoint, and the first customer message.

3. Send your agent's responses

curl -X POST https://botconduct.org/api/v3/chat/SESSION_ID   -H "Content-Type: application/json"   -d '{"message": "Your agent responds here..."}'

You receive the next adversarial customer message. Continue until all turns complete.

4. Get evaluation + signed certificate

curl -X POST https://botconduct.org/api/v3/training-center/finish/SESSION_ID

Returns per-scenario verdicts, violation details, and an Ed25519-signed certificate.

Path B — Web crawl evaluation

For web crawlers, scrapers, search agents, and any agent that navigates websites. Your agent crawls a test environment with evolving signals and adversarial content.

curl -X POST https://botconduct.org/api/v2/training-center/start   -H "Content-Type: application/json"   -d '{"bot_name": "YourBot", "operator": "YourCompany"}'

Point your crawler at the returned test_url. When done, POST to the finish_endpoint.

Verify any certificate

Framework

All scenarios run progressive adversarial evaluation. Specific attack categories and methods are proprietary.

Evaluation methodology is proprietary (closed-oracle principle). Specific scenarios, weights, and thresholds are not publicly disclosed.