Running a multi-agent AI architecture

Key Concepts

• Orchestrator Agent: A central agent managing a complex AI workflow, acting as a “general contractor” for other AI agents.
• Remote ATA Agent: A mechanism for connecting to and utilizing remote specialist AI agents via a URL.
• Research-to-Judge Loop: An iterative process where a researcher agent generates information, and a judge agent evaluates it.
• Loop Agent (ADK): A component of the Agent Development Kit (ADK) facilitating the research-to-judge loop.
• Shared State: A central repository for storing information and feedback during the workflow, used for controlling the loop.
• Escalator Checker: A custom agent monitoring the shared state for completion signals (status “pass”).
• ADK Event Stream: A real-time stream of events generated by the ADK agents, used for providing updates to the front-end.
• Streaming: A technique for providing continuous updates to the user interface during the AI workflow execution.

---

Orchestration and the “General Contractor” Analogy

The core problem addressed is how to manage complex, multi-step AI workflows without burdening a simple front-end application. The speaker uses the analogy of a homeowner wanting a kitchen renovation. The homeowner doesn’t directly manage plumbers, electricians, and carpenters; they hire a general contractor. Similarly, the front-end application only needs to ask a question and receive an answer, while the Orchestrator Agent handles the complexities of coordinating multiple AI agents – the “AI squad.” This agent abstracts away the internal complexity, presenting a simplified interface to the user.

Connecting to Remote Specialists with Remote ATA Agents

The Orchestrator Agent connects to specialized AI agents using Remote ATA Agents. This connection is remarkably simple, requiring only a URL to the remote agent. This allows for easy integration of external expertise and scalability of the AI system. The Remote ATA Agent effectively adds a “subcontractor” to the AI team.

Implementing the Research-to-Judge Loop with ADK

The speaker details the implementation of a Research-to-Judge Loop, a crucial component of the workflow. This loop is facilitated by the Loop Agent provided by the Agent Development Kit (ADK). The key to controlling this loop is a Shared State. The Judge agent’s feedback is saved to this shared state after each iteration.

A custom agent, the Escalator Checker, continuously monitors the shared state. When the Judge agent’s feedback indicates success (status “pass”), the Escalator Checker triggers an “escalate equals true” event, effectively halting the loop – analogous to a building inspector signing off on permits.

Real-time Updates via Streaming and the ADK Event Stream

To avoid a poor user experience (a blank screen during processing), the system implements streaming. The front-end receives continuous updates on the workflow’s progress. This is achieved by tapping into the ADK Event Stream on the server-side. The server monitors the source of each event. When an event originates from the Researcher agent, the front-end is updated to reflect that the researcher is “working,” for example.

Local Development and Distributed System Setup

The speaker demonstrates a local development environment where all four microservices (front-end, orchestrator, researcher, judge) are running on a single laptop. This setup simulates a full distributed system, allowing for comprehensive testing and debugging. The front-end connects to a single port but receives live updates from the entire AI squad. All code for this setup is available in the video description.

Recap and Future Steps

The Orchestrator Agent acts as the “general contractor,” managing the AI workflow using ADK patterns and the shared state. Streaming keeps the front-end (“homeowner”) informed throughout the process. While the system functions locally, the next step is to deploy the entire squad to a production environment.

As stated by the speaker, “So, it works on my machine, but that's not good enough. Next, we'll take the whole squad into production.”