Create agent

Agent behavioral config. The keys depend on type.

type=agent, built inside Calibrate:

• system_prompt: the agent's instructions
• llm.model: provider/model, e.g. openai/gpt-4.1 or google/gemini-2.5-flash
• stt.provider: deepgram, openai, cartesia, elevenlabs, google, sarvam, or smallest
• tts.provider: cartesia, openai, google, elevenlabs, sarvam, or smallest
• settings.agent_speaks_first, settings.max_assistant_turns
• system_tools.end_call: let the agent end the call
• data_extraction_fields: [{name, type, description, required}]

{
  "system_prompt": "You are a helpful support agent.",
  "llm": {"model": "openai/gpt-4.1"},
  "stt": {"provider": "deepgram"},
  "tts": {"provider": "elevenlabs"},
  "settings": {"agent_speaks_first": true, "max_assistant_turns": 50}
}

type=connection, your own HTTP endpoint:

• agent_url: public HTTP(S) endpoint your agent is called at
• agent_headers: headers sent on each request, e.g. auth
• benchmark_provider: openrouter by default. Other values: openai, google, anthropic, meta-llama, mistralai, deepseek, x-ai, cohere, qwen, or ai21

{
  "agent_url": "https://api.example.com/agent",
  "agent_headers": {"Authorization": "Bearer <token>"},
  "benchmark_provider": "openrouter"
}

For type=agent, omitted keys inherit managed defaults. Omit config entirely to use all defaults. For type=connection, config is stored as-is and must contain agent_url