DialogLab is a research prototype that provides a unified interface to configure conversational scenes, define agent personas, manage group structures, specify turn-taking rules, and orchestrate transitions between scripted narratives and improvisation. Designers can 1) configure group, party, snippet characteristics, 2) test with simulation and live interaction, and 3) gain insights with timeline view and post-hoc analytics.

KiloClaw – OpenClaw in under 60 seconds — Choose from 500+ models and get 7 days of free compute

Choose from 500+ models and get 7 days of free compute

Every AI product built in the last 3 years was optimized for the same interaction: one human, one AI, one thread. But that's not how real conversations work.

Team standups. Classroom discussions. Conference Q&As. Game NPC interactions. These are multi-party, fluid, and full of interruptions.

We've been designing AI for the exception, not the rule.

DialogLab is Google Research's open-source framework to author, simulate, and test dynamic group conversations involving both humans and AI agents.

Not just "prompt a model and hope for the best." An actual design environment for multi-party dialogue.

Here's what it does differently 👇

Visual scene builder - drag-and-drop canvas to set up participants, roles, subgroups, and shared content
Snippet-based flow control - break conversations into phases like opening, debate, and consensus, each with its own turn-taking and interruption rules
Human-in-the-loop simulation - an audit panel surfaces AI response suggestions during testing; you accept, edit, or dismiss in real time
Verification dashboard - visualizes turn distribution and sentiment flow so you're not reading 200 lines of raw transcript

Tested with 14 domain experts across game design, education, and social science research. Human control mode rated consistently higher on engagement, realism, and effectiveness vs fully autonomous agents.

Who should care right now 🎯

Game devs building NPC dialogue systems that feel alive
Educators creating AI-simulated practice environments like mock interviews or debate prep
Researchers studying group dynamics without coordinating 10 humans in a lab
Product teams prototyping AI experiences beyond the single chatbot window

We're entering a world where AI agents talk to each other and to humans simultaneously.

DialogLab is an early, honest attempt to build it. It's a research prototype, not a polished SaaS tool. The GitHub is open, the paper is published. Worth an afternoon if you're building anything in the multi-agent or conversational AI space.

GitHub: https://github.com/ecruhue/DialogLab

If you could simulate any group conversation before shipping it, what would you test first? Share in the comments! :)

Hey Rohan, that line about AI being optimized for the exception, not the rule, is a sharp reframe. Was there a specific project or use case where you watched a single-thread AI completely fail in a group setting?

This is interesting, I would like to see multi party chats that include humans and their AI assistants.

congrats @sundar_pichai for the launch