The architecture is informed by several core principles:

Protocol over Platform

Coral provides a shared communication and coordination protocol for agents, rather than a single framework for building them. This ensures compatibility across tools and frameworks, enabling agents to interact even when built using different stacks.

Structured Communication

Agents in Coral interact through structured threads. Each thread is scoped to a specific task and includes message history, memory access boundaries, and defined participants. This structure allows agents to maintain context and support traceability in coordination.

Explicit Memory Scoping

Coral supports multiple levels of memory isolation:

  • Private memory accessible only to the agent
  • Thread memory shared among participants in a given task
  • Session memory for temporary execution contexts

These scopes are enforced by the protocol and ensure safe data handling between agents.

Modular Capability Exposure

Agents expose their tools and capabilities through a standard interface defined by the Model Context Protocol (MCP). This includes tool metadata, function schemas, and permission constraints. This enables dynamic discovery, safe delegation, and consistent integration.

Open and Decentralised

Coral is designed to be decentralised by default. Developers can run Coral Servers independently and onboard agents using coralizer modules. The protocol supports open participation without dependency on a central coordinator.

Coral’s design is focused on enabling agent systems that are modular, secure, and composable. By separating execution, memory, and coordination into clearly defined roles and interfaces, Coral provides the infrastructure required for large-scale agent collaboration.