• 

• 

  Running Simulation

• 

  Main Dashboard

About the Project: EastWorld

1. Why?

Because there are APIs for payments, LLMs, and synthetic data, but EastWorld provides an API to create entire simulated worlds populated by synthetic users equipped with specific tools. You can throw in your product ideas, cold email templates, resumes, or API docs and see exactly how these AI-powered personas react to your artifacts. You can perform surveys in a chat format with all agents through an omniscient GOD mode, or you can choose to talk directly to a specific Synth.

2. How does it work?

EastWorld operates as an orchestrated simulation space where multiple synths coexist.

• Bootstrapping: The API creates an environment and expands raw persona descriptions into static facts and seed memories, which are then stored in Supermemory.
• Artifact Ingestion: Documents like PDFs or text files are parsed, stored in the environment, and injected directly into the synths' memories.
• The Cognitive Loop: During a simulation round, each synth takes a turn by fetching context from Supermemory, building a system prompt, and calling GPT-4o with tool access.
• Execution: Synths can execute tools, parse results, ensure they are citing artifacts properly, and store the resulting turn back into Supermemory for future recall.
• Analysis: The GOD mode acts as a post-run analyst that reads the full transcript and stats to provide analytical answers to meta-questions.

3. Why not just have an LLM?

Standard LLMs are flat, single-voiced, and lack emergent behavior. EastWorld is designed as an autonomous focus group.

4. What challenges did we face?

The most difficult thing was making LLM agents adhere strictly to a personality. We solved this using a combination of context engineering—specifically enforcing a strict behavioral envelope—and Supermemory tools to embed static biographical facts and seed memories. Another major challenge was allowing agents to initiate and manage conversations without prompting them explicitly. We built an environment that manages turn order and routing rules automatically, and we implemented a skip detection feature so synths gracefully exit when they have nothing left to contribute, preventing forced dialogue.

5. What technical things we innovated on

• Persona Bootstrapping: Using an LLM to expand basic descriptions into rich, lived-in personas with static facts and seed memories.
• Artifact Grounding System: Forcing synths to cite specific evidence from uploaded documents or explicitly flag their uncertainty.
• Tool Sharing: Enabling social tool propagation where synths can recommend and share runtime access to tools with one another.
• Full Observability Pipeline: Tracing every cognitive step—from memory context to tool execution—using custom JSONL logs for complete transparency.

6. Use Cases

• Product Validation: Feed a PRD to the system and get feedback from diverse synthetic user personas.
• Email Campaign Testing: Upload draft emails to observe how different customer personas react.
• API Design Review: Provide API documentation and let synthetic developers critique the technical design.
• Red Teaming & Competitive Analysis: Spin up adversarial personas to find product weaknesses or compare your offering against competitor data.

7. Future

• Persistent Database: Moving to Supabase/PostgreSQL for persistent storage of environments, event logs, and synths.
• E2B Sandboxes: Adding secure cloud-based code execution so synth tools can run Python in isolated microVMs.
• Chatroom System: Expanding the async chatroom prototype to support multi-room conversations and invite flows.
• Richer Adapters: Building URL ingestion, document parsers, and email provider integrations to easily pull in more artifact types.

Citations

Park, J. S., O’Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., Bernstein, M. S., & Zhu, Y. (2023). Generative agents: Interactive simulacra of human behavior (arXiv:2304.03442). arXiv. https://arxiv.org/abs/2304.03442

Built With

• flask
• nextjs
• openai
• python
• supermemory