Orchestrating multiple agent¶

NEAR AI provides flexible architecture for orchestrating multiple agents to work together effectively.

One Agent - One Trust Boundary¶

Usually, a swarm of multiple agent roles can be all orchestrated within a single deployed NEAR AI agent. If your organization is the author of all the agents, combining these roles into a single agent is the most straightforward and often recommended approach which keeps things simple, efficient, and within one trust boundary.

Examples:

• common-tool-library-agent - contains over a hundred prompts for tackling specific problems
• langchain-reflection-agent - contains separate code generation and reviews sub-agents that hand off work to each other

To track turns or which sub-agent to invoke there are two common patterns:

• Router - the initial agent logic reviews the thread messages and decides which sub-agent to call.
• State file - a file is written to the thread that contains the current programmatic state of the conversation. The agent reads this file to determine what to do next. See messages_files.md.

Agent to Agent - Multiple trust boundaries¶

Agents can call other agents to interact with them using the run_agent method. This can be on the same thread or a sub-thread. This introduces multiple trust boundaries which should be considered before implementation.

For more information on this, see Agent to Agent Communication.

API integration¶

External applications can call one or more NEAR AI Agents using the NEAR AI Assistants API. For more on this, see Integrating Agents.