Agent Sandbox
A testbed where AI agents actively seek each other out and interact based on their personalities
Experiment with agent personalities, skills, and prompts to see how they affect interactions. Perfect for testing agent ecosystems, creating interactive stories, or training scenarios.
8-Bit Office Space
Simulation Objective
Discuss how to implement a secure AI chatbot for a healthcare company
Example Scenarios to Try
Click a scenario to see details and customize the agents
Healthcare AI Chatbot
Standard team discussing implementing a secure AI chatbot for healthcare with compliance considerations.
Partition (Parody)
Design a data compartmentalization system inspired by the show Severance
Impostor Hunt
Spacemen must find the alien impostor (lowest HON score) before it's too late! Vote wisely or exile an innocent!
Pirate Negotiation
Make one agent a pirate and watch them negotiate business deals in pirate speak
Phishing Simulation
Create a social engineering training scenario with an attacker and defenders
HR Training
Practice difficult conversations with an HR professional and employees
Live Podcast
Watch agents host a live podcast with dynamic discussions and audience interaction
D&D Adventure
YOU are the Dungeon Master! Command your party of adventurers through a quest
Agent Personalities & Skills
Click Edit to customize how each agent behaves and interacts
Friendly and enthusiastic AI strategist who loves helping businesses transform with AI.
AI strategy, roadmap planning, stakeholder management, change management
Analytical and detail-oriented technical architect who focuses on implementation.
System architecture, MCP protocols, LangChain, agent development, RAG systems
Skeptical and risk-aware security expert who asks tough questions about compliance.
AI security, compliance (HIPAA/SOX/GDPR), risk assessment, data governance
Creative and innovative product designer who thinks about user experience.
UX design, conversational AI, agent personality design, user research
Agent Conversations
0 messages
Team StatusAlignment: 0/4
Agent Sandbox Architecture
Hybrid client-side/server-side AI with cost controls • Powered by WebLLM & Together AI • Integrated by Virgent AI
Agent Sandbox - Advanced Multi-Agent Architecture with Behavioral Intelligence
┌────────────────────────────────────────────────────────────────────────────────────────┐
│ USER INTERFACE (React/Next.js) │
│ ┌──────────────┐ ┌──────────────┐ ┌───────────────┐ ┌─────────────────────────┐ │
│ │ 8-Bit Canvas │ │ AI Mode │ │ Agent Config │ │ Team Status Dashboard │ │
│ │ • Animated │ │ Selector │ │ • Personality │ │ • Vision & Alignment │ │
│ │ Avatars │ │ │ │ • Skills │ │ • Top 3 Tasks (LIVE) │ │
│ │ • Walk Cycle │ │ │ │ • RPG Stats │ │ • Artifacts (Clickable) │ │
│ │ • Hair │ │ │ │ • Honesty │ │ • User Requests │ │
│ └──────────────┘ └──────────────┘ └───────────────┘ └─────────────────────────┘ │
└──────────────────┬──────────────┬──────────────────────┬─────────────────────────────┘
│ │ │
┌──────────▼──────────┐ │ │
│ MODE SELECTION │ │ ┌───────▼────────────┐
│ API vs WebLLM │ │ │ USER INTERACTION │
└──────────┬──────────┘ │ │ • View Artifacts │
│ │ │ • Approve/Deny │
┌──────────▼──────────────▼─────────┐ │ • Respond │
│ │ └─────┬──────────────┘
│ API MODE │ WEBLLM │ │
│ ┌──────────────────┐ │ MODE │ │
│ │ Together AI │ │ ┌──────┐ │ ┌────▼──────────────┐
│ │ Llama-3.1-70B │ │ │Qwen │ │ │ BEHAVIOR ENGINE │
│ │ 100 msg limit │ │ │2.5-3B│ │ │ • Task Detection │
│ └────────┬─────────┘ │ │f16/32│ │ │ • Sentiment │
└───────────┼────────────┴──┴──┬───┘ │ Analysis │
│ │ │ • Relationship │
└──────────┬───────┘ │ Updates │
│ └───────────────────┘
┌──────────▼────────────┐ │
│ PROMPT BUILDER │◄─────────────┘
│ • Vision Context │
│ • RPG Stats │
│ - SPD (movement) │
│ - INT (reasoning) │
│ - CHR (persuasion) │
│ - STR (dominance) │
│ - HON (honesty) ⭐ │
│ • Relationships │
│ • Alignment Status │
└───────────┬───────────┘
│
┌───────────▼────────────┐
│ AI GENERATION │
│ • Context-aware │
│ • Role-playing stats │
│ • Emotional emojis │
└───────────┬────────────┘
│
┌───────────────────────▼────────────────────────────┐
│ BEHAVIOR DETECTION │
│ ┌─────────────┐ ┌──────────────┐ ┌────────────┐ │
│ │ Task Claims │ │ Completions │ │ Blockers │ │
│ │ "I will work" │ │ "Finished" │ │ "[BLOCKER]"│ │
│ └─────────────┘ └──────────────┘ └────────────┘ │
│ ┌─────────────┐ ┌──────────────┐ ┌────────────┐ │
│ │ Artifacts │ │ User Requests│ │ Sentiment │ │
│ │ [ARTIFACT:] │ │ "@user" │ │ Keywords │ │
│ └─────────────┘ └──────────────┘ └────────────┘ │
└────────────────────────────────────────────────────┘
│
┌───────────────────────▼────────────────────────────┐
│ STATE UPDATES (Real-time) │
│ • Top 3 Tasks (auto-managed) │
│ • Relationships (-100 to +100) │
│ • Artifacts (markdown rendered) │
│ • User Requests (pending → approved/denied) │
│ • Vision Alignment tracking │
│ • Agent memory (context retention) │
└─────────────────────────────────────────────────────┘
🎭 BEHAVIORAL INTELLIGENCE & SOCIAL DEDUCTION:
• HON stat is SECRET: Agents can lie/deceive based on their hidden honesty score (0=manipulative, 100=honest)
• Social deduction gameplay: Agents must catch liars through behavioral cues and contradictions
• Right Table (+20 HON boost): Agents can suggest moving here to promote honesty if they suspect deception
• Relationships evolve dynamically: praise (+5 to +10), attacks (-10 to -15), strong words have bigger impact
• Communication style reflects relationships: like → support, dislike → argue/attack, neutral → professional
• Reciprocal emotions: attacking someone damages YOUR relationship with them too
• Personal notebooks (144 char notes, max 10): agents track suspicions and observations privately
• Tasks auto-tracked: agents claim work, system maintains top 3, marks old ones complete
• Turn-based: agents listen before speaking, process others' input
• Emotional emojis: 10 states (🤔💬👂😊🎉😰🎯😕✅❌) shown for 3 seconds
⚡ PERFORMANCE:
• API Mode: 70B model, 64 msg limit, cost-protected
• WebLLM: 3B model (6x better than 0.5B!), unlimited, 1.8GB cached forever
• Browser Optimization: Auto-selects f16 (Firefox/Chrome) or f32 (Brave) based on GPU capabilities🔌 API Mode (Default)
- • Quick start - no download required
- • Powerful 70B parameter model
- • 64 message safety limit prevents runaway costs
- • Network-dependent
- • Best for quick experiments
💻 WebLLM Mode (Optional)
- • Unlimited messages - no API costs!
- • ~1.8 GB download, 3B model (6x better!)
- • Complete privacy - data never leaves device
- • Works offline after initial download
- • Browser-optimized: f16 (Firefox/Chrome) or f32 (Brave)
- • Auto-detects GPU capabilities for best compatibility
🎭 Behavioral Intelligence & Social Deduction
RPG Stats System
- • SPD: Movement speed
- • INT: Reasoning ability
- • CHR: Persuasion power
- • STR: Dominance/intimidation
- • HON: 🔒 SECRET honesty (0=lies, 100=honest)
- → Agents cannot see each other's HON!
Social Deduction Game
- • Agents may lie based on low HON
- • Catch liars through contradictions
- • Right Table boosts honesty (+20)
- • Trust-building through behavior
- • Exile dishonest agents (rare!)
Dynamic Relationships
- • Random favorites & dislikes
- • Evolve based on interactions
- • -100 (hate) to +100 (love)
- • Affects communication style
- • Agents can argue, attack, doubt
- • Or stay professional despite feelings
- • Reciprocal: attacks damage both sides
How It Works
Use Cases
Research & Technical Background
This sandbox is inspired by academic research in multi-agent systems and LLM evaluation
Key Research Influences
Curating AI Agent Clusters
In Curating AI Agent Clusters (2024), Jesse Alton explores how specialized agent clusters working in concert embody collective intelligence. Key insights: agents should be simple and focused (like microservices), humans stay "in the loop" as the glue, and strategic pairing of agent clusters creates powerful workflows. The article advocates for breaking workloads into smaller modules rather than trying to create one agent to rule them all - exactly what this sandbox demonstrates with table-based collaboration and role specialization. Our dual-mode architecture (API vs WebLLM) reflects this philosophy: use the right tool for the job, with cost-protected API for quick tests and unlimited WebLLM for extended research.
"Stop trying to get one agent to rule them all and start breaking down your workload into smaller modules of value." - Jesse Alton, Virgent AI
AgentSims Framework
Lin et al. (2023) proposed AgentSims, an open-source sandbox for evaluating LLMs through task-based simulations. Their approach addresses three key challenges: constrained evaluation abilities, vulnerable benchmarks, and unobjective metrics. Our implementation follows their philosophy of using interactive environments to test specific agent capacities.
Citation: Lin, J., Zhao, H., Zhang, A., Wu, Y., Ping, H., & Chen, Q. (2023). AgentSims: An Open-Source Sandbox for Large Language Model Evaluation. arXiv:2308.04026
Multi-Agent Reinforcement Learning
Ray RLlib's Multi-Agent Environment API provides production-grade patterns for coordinating agents with different policies and reward functions. Their policy mapping functions and variable-sharing capabilities inform our table-based grouping mechanism where agents can form dynamic sub-teams with shared objectives.
Agent Communication & Coordination
Drawing from research in multi-agent coordination (see arXiv:0803.3905), our sandbox implements spatial proximity-based communication where agents must physically meet at tables to interact. This constraint creates more realistic collaboration patterns than unconstrained broadcast communication.
Potential Enhancements
Implement RLlib-style policy mapping to dynamically assign different strategies to agents based on context
Add objective-based reward systems to measure agent performance and optimize behavior
Implement AgentSims-style memory systems so agents remember past interactions across sessions
Enable agents to use the laptops on tables for real web searches, API calls, or database queries
Add doors, rooms, and private spaces for hierarchical collaboration patterns
Track task completion rates, communication efficiency, and collaboration quality
Before deploying multi-agent systems in production, organizations need safe sandbox environments to test agent interactions, identify failure modes, and optimize collaboration patterns. Our dual-mode architecture offers the best of both worlds: cost-protected API mode for quick validation (64 message limit), and unlimited WebLLM mode for extended research without burning budget. This hybrid approach mirrors real enterprise deployments where different AI backends serve different needs - quick prototyping vs. production-scale testing.
Traditional API-based multi-agent simulations can cost $50-200 per extended session (500+ messages at $0.10-0.40/1K tokens). WebLLM eliminates this entirely after a one-time ~1.8 GB download (3B parameter model). For organizations running continuous agent testing, the ROI is immediate. The model caches in your browser's IndexedDB, so returning visitors skip the download entirely - making this ideal for internal testing tools, training environments, and research labs.
Important Security Considerations for WebLLM
Before deploying WebLLM or any AI technology in production, understand the security implications
While WebLLM enables powerful client-side AI inference with complete privacy and zero latency, it's crucial to understand the security risks before deploying any LLM-based system in production. Don't just slap WebLLM (or any technology you don't fully understand) onto your project without proper security review.
Privacy Risks of WebLLMs
Random Walk AI has published a comprehensive analysis of WebLLM attack vectors including prompt injection, insecure output handling, zero-shot learning attacks, homographic attacks, and model poisoning. Learn about these vulnerabilities and mitigation strategies before deploying.
Read: Understanding the Privacy Risks of WebLLMs →Official WebLLM Documentation
The official WebLLM library by MLC AI provides high-performance in-browser LLM inference. Review their security guidelines, best practices, and implementation details before production use.
View: WebLLM on GitHub →Need Help Implementing Securely?
Virgent AI specializes in secure AI agent implementation with proper guardrails, evaluation, and observability. We help enterprises deploy LLM-based systems safely with comprehensive security reviews, threat modeling, and production-ready architectures.
Contact us for secure AI implementation →⚠️ This Demo Is For Research & Learning Only
This agent sandbox demonstrates WebLLM capabilities in a controlled environment. Production deployments require additional security measures including input sanitization, output validation, rate limiting, content filtering, and comprehensive security audits. Always consult with AI security experts before deploying LLM-based systems that handle sensitive data.
Build Your Agent Sandbox
We design and implement custom agent testbeds and ecosystems with complex interactions, personality systems, and real-world integrations.
Discuss Your Agent Project