MAESTRO lets you design synthetic social systems: a human participant interacts in real time with several LLM-based actors, each with its own personality and communication style. You draft the conversation and your experimental condition in advance, and an orchestrating LLM, the conductor, keeps the interaction coherent whenever the participant writes something. Every participant gets a standardized yet responsive manipulation inside a conversation that reacts to the participant.
Field studies capture social dynamics as they really unfold but offer little control, and laboratory studies control the situation but rarely feel real. MAESTRO lets you set what the actors do and when, keeps the manipulation consistent across participants, and still responds naturally to what each participant says. MAESTRO allows insights into aspects that were hard to examine before: rare social conditions, micro processes as they unfold, and reciprocal dynamics between actors.
The participant joins a synthetic social system of LLM-based actors who interact with the participant and with each other. You can give each actor its own personality, demographic characteristics and communication style. By default the actors appear as colors (Blue, Green, Red) to avoid priming participants in terms of gender, age and ethnicity. You can also give the actors names that fit your study.
The reference script is a chronological template of the actors' messages. Each message has a sender, a send time and a typing window, so you design the rhythm of the conversation: quick replies, long silences, slow typers. Every participant in a condition starts from the same script, which keeps the timing comparable across sessions.
Whenever the participant sends a message, the conductor rewrites the actors' not-yet-sent messages so they respond to the prior interaction. The conductor follows your instructions (task, personalities, condition) and never appears in the chat itself. Participants feel heard, and your manipulation stays consistent.
You can divide the reference script into interaction segments. Each segment can introduce a targeted change at a set time, for example new information that arrives 4 minutes into the conversation or an actor who abruptly leaves the conversation, so the participant's immediate response becomes observable.
You can ask the assistant to draft a conversation from your task description, or write "make Green more skeptical" and watch the assistant rewrite the affected messages in front of you, with an undo for every change. The assistant can help with the conductor instructions in the same way, and you can still edit everything by hand.
Every message is stored with millisecond timestamps, sender and conversation ID, so response times and turn taking can be analyzed. You can export CSV for common statistical software, PDF transcripts for the appendix, or one ZIP with everything. Test sessions stay tagged, so pilot runs never mix with real data.
A fixed script sounds robotic the moment a participant says something unexpected. So an orchestrating LLM, the conductor, sits above the actors. Whenever the participant sends a message, the conductor analyzes that message, the preceding conversation and the reference script, and cues the actors so that pacing, turn taking, personalities and your experimental condition stay coherent. The conductor stays invisible to the participant and never writes in the chat itself. One full cycle has five steps:
Blue, Green and Red send their messages at the times you scheduled, and a typing indicator appears before each message.
Yellow contributes an idea, an objection or a question.
The conversation so far, the remaining script and your instructions are sent to the conductor, who returns an updated version of the remaining conversation that responds to Yellow. Personalities, pacing and condition stay as defined.
The exchange keeps its flow: the next actor simply types a reply that fits what Yellow said, at the pace you defined.
To the participant, the conversation is a single natural group chat. To you, the conversation is a session that followed your manipulation and is comparable across your sample.
Everything runs in the browser, for you and for your participants. The studio includes the example studies from our paper, which you can try in a test chat and copy as a starting point for your study. You can share a study setup so that other researchers can inspect, reuse and replicate the study.
You describe your study and decide what participants see: task instructions above the chat, a waiting screen while the other actors join, a "ready?" check and a session timer.
You can choose between OpenAI, Anthropic or other LLM gateways (for example your university's gateway), pick the model and set its thinking effort. MAESTRO estimates the API costs. API keys stay on the server.
You name the actors and give the actors personalities, write the conductor instructions for your condition, lay out the reference script and, if you like, add segments where the situation should change. The AI assistant can help with the actors, the conductor instructions and the reference script.
You can run test sessions yourself before you put the participant link into Qualtrics or Prolific with a piped conversation ID. Sessions arrive on the data page while the study runs, and you can export the sessions whenever you like.
Participants arrive from your survey, chat with the actors, and return to the survey with the conversation ID attached.
The assistant can help you write both the reference script and the conductor instructions. You can describe what you want in plain language, for example "make Green more skeptical", and the assistant rewrites the affected messages right in the editor and lists the changes. Every change can be undone.
Anonymized for peer review
Anonymized for peer review
Anonymized for peer review
Anonymized for peer review
The paper is under peer review, so the citation is anonymous for now. The author names will appear here once the paper is accepted for publication.
MAESTRO is free and open source. You can try an example study in a test chat, copy the example study and adapt the copy to your research. A first test session takes only a few minutes.
Open the studio