Building Multi-Player AI Systems (and why it’s SO hard)
MeshAgent introduces a multiplayer AI paradigm, shifting from single-user systems to collaborative 'Rooms' where teams of humans and agents can work together with shared context. This talk explores the platform's architecture, developer tools, and its approach to solving real-world collaborative tasks.
Most AI development has focused on a single-player paradigm: one human interacting with one or more agents. However, real-world business processes are a team sport, involving multiple people and systems working towards a shared goal. MeshAgent addresses this gap by providing an end-to-end platform to build, deploy, and operate AI agents in a collaborative, multiplayer environment.
The Shift to Multiplayer AI with Secure Rooms
The core problem with existing agent frameworks is their failure to address what happens when an entire team needs to interact with AI agents to build something together. MeshAgent’s solution is centered on the concept of a Room.
A Room is a secure, on-demand workspace that acts as a security boundary, uniting teams of humans, agents, tools, and data sources. Key features of a Room include:
- Shared Context: All participants in a Room, whether human or agent, share the same context, ensuring everyone is working with the same information and deliverables.
- Unified Communication: The underlying infrastructure is designed to handle communication seamlessly between all participant types. Humans can message other humans or agents using the same methods, treating agents as true teammates.
- Dynamic Infrastructure: A Room automatically spins up the moment the first participant joins and spins down when the last one leaves. This ensures that infrastructure is only running when needed. Artifacts and documents created within the Room can be saved to external storage for persistence.
A Practical Use Case: Health Insurance RFP Response
To illustrate the concept, consider a health insurance company responding to a Request for Proposal (RFP). This process requires collaboration between multiple departments like underwriting, actuarial, sales, and legal.
With MeshAgent, a dedicated Room can be created for each RFP.
- Participants: Members from each human team can be invited into the Room.
- Specialized Agents: Each team can bring in their specialized agents. For example, the actuarial team might use an agent equipped with a tool to call a custom AI model for risk or liability analysis. Other agents could be responsible for writing and assembling the final document.
- Seamless Collaboration: The human team can enter and exit the Room as needed to check progress. If an agent requires human input, it can send a notification to the relevant person, enabling a fluid workflow between all parties.
- Scalability: This model scales easily, allowing the company to spin up a new, isolated Room for every RFP they need to handle, ensuring data is properly scoped and secure for each client.
Developer Platform and Observability
MeshAgent is designed to be a comprehensive platform for the entire agent lifecycle, from building to operations.
- Multi-language SDKs: To ensure developers can work in their preferred environments, MeshAgent provides SDKs in Python, TypeScript/JavaScript, and Dart.
- MeshAgent Studio: For the building and testing phase, the Studio provides an environment to evaluate agent performance. It allows developers to see all participants in a room, inspect the tools available to agents, and interact directly with them via chat. Shared documents and artifacts appear automatically, providing a clear view of the collaborative workspace.
- Built-in Logging and Observability: The platform has logging and observability baked in. It automatically tracks agent activities, including which tools were called, messages sent, and execution times. This creates a complete audit trail, ensuring transparency and trustworthiness in the final output. The system integrates with Open Telemetry and is customizable to incorporate other observability tools for a “single pane of glass” view for operations teams.
Platform Flexibility
During the Q&A, key aspects of the platform’s flexibility were highlighted:
- Scalability: There are no restrictions on the number of Rooms. The system can scale to handle thousands of concurrent rooms, spinning up and down as needed.
- Configurable Memory: Developers have full control over how agent memory is handled. It can be configured to be persistent across sessions by saving and loading state, or it can be flushed as required by the specific application, all while respecting the underlying model’s context window constraints.