From: aidotengineer
Multiagent systems involve multiple AI agents interacting with each other [01:01:10]. This contrasts with traditional AI agent architectures where agents might be trapped within a single Python file, unable to run asynchronously or exist independently [01:01:31] [01:01:33].
Limitations of Current Multiagent Frameworks
Many existing multi-agent frameworks, such as Autogen, often lack true independence among agents [01:02:23] [01:02:28]. These agents are typically stateless, meaning they do not retain experiences or skills over time, limiting the benefits of a multi-agent setup [01:02:05] [01:02:07]. For instance, an expert agent from one multi-agent group cannot be easily extracted and used in another [01:02:11].
Leta’s Approach to Multiagent Communication
Leta addresses these limitations by enabling stateful agents that run as services backed by APIs [01:02:17] [01:02:20]. This design means that multi-agent interaction is fundamentally achieved through message passing over APIs [01:02:24] [01:02:28]. The ability to import the Leta client directly within a tool allows agents to send messages to other agents, create new ones, and manage their memory [01:02:51] [01:02:57].
Types of Multiagent Communication
Leta supports different patterns for multi-agent communication:
- Asynchronous Messaging [01:03:11]: This mimics human communication, where an agent sends a message and immediately receives a receipt, allowing it to continue its execution without waiting for a reply [01:03:17] [01:03:22].
- Synchronous Messaging [01:04:00]: In certain scenarios, it’s beneficial for an agent’s execution to pause until a reply is received, for example, when an agent needs to consult a supervisor [01:03:49] [01:03:54].
- Group Messaging (Supervisor-Worker Concept) [01:04:09]: Agents can be grouped using tags, allowing a supervisor agent to broadcast messages to all agents matching specific tags, facilitating tasks like delegated research or parallelized operations [01:04:10] [01:04:29].
Implementation and Benefits
The multi-agent messaging tool is crucial for enabling communication between agents in Leta [01:07:15] [01:07:19]. This tool, like others in Leta, operates within a sandbox environment, preventing interference between different agents’ tools [01:00:16] [01:00:30].
The ability for agents to be stateful and communicate via APIs unlocks significant potential for multiagent orchestration, allowing for human-like interaction patterns [01:03:09]. Agents can be easily managed by adding or removing tools, which can immediately impact their ability to communicate or perform specific actions [01:11:37] [01:11:40].
A prominent use case for multi-agent communication systems is in enterprise deployments for stateful workflows where agents process transactions and learn about users over time, without necessarily relying on a chatbot interface [01:13:21] [01:13:33].