Carrier Log in
← All tutorials

AI agentsCommunicatorChat

An AI partner in a chat

Make an agent reply with an AI model. Set its provider, model, prompt and reply timing. Add the experiment's API key and change the provider.

4 steps 6 min read

  1. LLM Response

    Click the agent in the chamber's cast. In its settings, open LLM Response. When you complete and save this section, the agent replies with that model. The header of the section then shows Enabled.

    Set these items:

    • The provider: OpenAI, Anthropic or Google.
    • The model.
    • The system prompt: who the agent is and how it replies. Here Mira is a resident who prefers more berths. She replies in one or two sentences.
    • Context Window and Include chat history: how much of the conversation the agent reads.
    • Idle Detection Time: how long the agent waits after the last message before it replies. The default is 4 seconds, so the agent replies one time to a group of fast messages. With 0, it replies to each message immediately.
    • Response Delay: a random pause between two numbers, in milliseconds. Thus the reply looks like typing.

    Click Deploy to save.

    Mira uses gpt-4o-mini from OpenAI, with a prompt for a resident and a delay of 1.5–2.5 seconds.

  2. The experiment's API key

    In the settings of the experiment, API Keys has a key for each provider.

    CAUTION: Carrier cannot find out if a person uses a key incorrectly. Before you save a key, set a spending limit on it in the dashboard of the provider. Revoke the key when the study ends.

    Carrier encrypts a saved key and never sends it back to the browser. The field shows only the first and last characters. The provider charges you for the use of the key.

    If the key for a provider is empty, Carrier uses its own key. The cost then comes from the Carrier credit of the experiment owner. The owner's dashboard shows this credit.

    An OpenAI key saved for this study. The field shows only sk-tu…0000.

  3. Switching provider

    To use a different model, select a different provider in LLM Response. The model changes to the default model of that provider. For Anthropic, the default is claude-sonnet-4-6. You can change it.

    The agent then needs a key for that provider in API Keys. If there is no key, it uses the Carrier credit of the owner.

    Click Deploy to save.

    Mira changes to Anthropic.

  4. In the chat

    Mira reads the conversation and replies as her character. She replies some seconds after the resident stops typing.

    The chat history in the data export contains each reply, with the provider and model that wrote it.

    A resident speaks for the ferry. Mira speaks for the berths.