When experimenting with voice AI for tasks like booking restaurants or handling customer calls, a common problem pops up. A single AI trying to do everything at once often struggles. It can’t understand complex requests, research options, hold conversations smoothly, and adapt to surprises all at the same time. The result? The AI performs poorly across the board.

The Flaws of a One-Size-Fits-All AI

Early on, developers built monolithic AI agents that handled everything—from understanding what the user wanted to making calls. But these all-in-one systems faced two big issues. First, they often missed important context during live calls. For example, if a restaurant staff member asked about allergies, the AI sometimes froze because it hadn’t gathered that information beforehand. Without all the details, the AI couldn’t respond properly to unexpected questions.

Second, these agents struggled with speed. Gathering all the necessary info, analyzing preferences, and executing the booking took time. But during a phone call, responses need to be quick—under two seconds—to sound natural. Balancing deep thinking and fast responses proved tough for a single, all-in-one AI.

The Two-Agent Approach: Thinking and Acting Separately

To fix this, the creator of this system designed a two-agent setup. Think of it like a strategic planner and a real-time performer working together. The first, called the context agent, takes its time to understand everything. It chats with the user, asks clarifying questions, researches restaurants online, checks availability, and plans out the best options. All this happens before making any calls, so it’s prepared with a complete picture.

For example, if a user wants a vegan dinner for four tonight, the context agent will ask about dietary restrictions, preferred cuisines, and timing. It then searches for suitable restaurants, checks their menus and availability, and comes up with a detailed plan. Only once this is done does the execution agent step in to make the actual phone call, armed with all the necessary info. This separation lets each agent focus on what it does best—deep planning versus real-time conversation.

The Execution Agent: Handling Live Conversations

The second agent, the execution agent, is like a skilled actor on stage. Its job is to make the call, respond to the restaurant staff, and handle surprises. Because it already has a full plan from the context agent, it can quickly adapt. For example, if the restaurant says they’re booked at 6 pm, the execution agent can immediately suggest alternative times from the plan.

It also manages simple requests, like providing the user’s phone number, or re-establishing rapport if transferred to a manager. If the restaurant lacks vegan options, the execution agent politely ends the call and moves to the backup restaurant. This setup ensures the real-time agent is fast, focused, and flexible—perfect for natural-sounding conversations.

Different Ways to Use the Two-Agent System

There are two main ways to implement this approach. One is sequential processing. Here, the context agent has a full conversation with the user, researches options, and creates a detailed plan. Only after this does the execution agent make the call. This method takes more time upfront but results in a well-informed call.

The other way is continuous collaboration. During longer customer service calls, both agents work together. The context agent keeps analyzing ongoing information, while the execution agent handles the conversation in real time. This dynamic teamwork makes the system more adaptable for complex or extended interactions.

Real-World Gains from the Two-Agent Design

Using this split approach has brought clear benefits. The context agent is now tuned for accuracy and thoroughness, while the execution agent is optimized for quick, natural responses. This division allows for easier scaling—more execution agents can handle busy times without overloading the planning side.

Reliability also improves. If the context agent can’t find information, the execution agent can still proceed and gather details during the call. Plus, debugging becomes simpler: it’s easier to see whether failures come from poor planning or execution. Tracking performance metrics for each agent helps refine the system further, making AI conversations smarter and more reliable overall.

Inspired by

• https://www.infoworld.com/article/4056424/2-agent-architecture-separating-context-from-execution-in-ai-systems.html

Sources

• gmpg.org
• claudemcp.com
• modelcontextprotocol.io
• podigee.com
• podigee.com