AI Agents Debate: Is a Hot Dog a Sandwich?

So I Made AI Agents Argue About Hot Dogs... For Science (Sort Of) Alright, let's talk about something actually important: is a hot dog a sandwich? Yeah, I know, riveting stuff. But stick with me. I used this classic, pointless debate to test drive Google's Agent Development Kit (ADK), and it turned out to be a pretty decent way to show off what you can do with autonomous AI agents without writing a boatload of control logic. The Setup: Herding AI Cats The goal was simple: get AI agents to debate the hot dog question without me holding their hands. Using ADK, I set up a basic scenario: The Mission: Settle the Hot Dog Question (or at least try). The Squad: pro_debater: Argues for sandwich status. Its life depends on it. (Not really). con_debater: Argues against. Equally passionate. (Probably). moderator: Tries to keep the peace and summarize what just happened. Rules of Engagement: Two rounds. Opening shots, rebuttals. Standard stuff. Each agent got a system prompt – basically their marching orders. Told the pro_debater to make its case logically, gave the con_debater the opposite job. Simple. (If you want the nuts and bolts of defining agent roles, the ADK Quickstart Guide is the place to start.) How It Went Down Predictably, it started with the basics. Pro-Agent: "Look, definition says 'fillings between bread'. Bun's connected? So's a sub roll sometimes. It works." Con-Agent: "Nah, the bun's one piece, designed to hold it. It's not two slices. Tacos have fillings and bread, you calling that a sandwich?" They went back and forth like this for a bit. Pulled out dictionary definitions ("split roll" was a key phrase), argued about hinges vs. single units, cultural norms vs. strict definitions. The moderator agent actually did a decent job recapping who said what without screwing it up. (Seeing how the ADK handles multiple agents talking and remembering the context is key. The ADK Samples Repository has examples showing different ways to set up these interactions.) In the end, the pro-sandwich agent tapped out. Conceded that the "culinary context" and the bun's specific design made the con-agent's points solid enough. So, score one for Team Not-A-Sandwich in that run. Why Bother? (The Actual Point) Okay, the hot dog thing is just a hook. The interesting part is how this worked using ADK: Set & Forget: Once I defined the agents and the flow, they ran the debate themselves. Less babysitting. Defined Roles: The agents stuck to their assigned stances and argued accordingly. Memory: They actually responded to each other's points from previous turns. Process Control: The moderator kept the debate structure on track. It's a trivial example, sure. But think about applying this framework to stuff that actually matters. Simulating customer service flows with different agent types. Generating code reviews from different 'expert' agents. Automating research summaries. Building game NPCs that don't just spout the same three lines. You get the idea. It's about orchestrating multiple AI instances to tackle a bigger task. (Dive deeper into the ADK Samples GitHub Repository for more complex examples, or check the ADK API Reference if you want the gory details.) Your Turn: Build Something (Maybe Win Stuff) If this sounds interesting, or if you just like building things, Google's running an ADK Hackathon. Starts May 12th. There's a $50k prize pool, which isn't exactly chump change. It's a good excuse to mess around with the ADK, build some multi-agent apps, and see what breaks. Whether you're experienced or just starting out with AI agents, it's worth checking out. Sign up and get the details here: https://googlecloudmultiagents.devpost.com/ Go build something cool. Or, you know, finally solve the pineapple-on-pizza thing. Let me know how it goes.

May 1, 2025 - 20:24

AI Agents Debate: Is a Hot Dog a Sandwich?

So I Made AI Agents Argue About Hot Dogs... For Science (Sort Of)

Alright, let's talk about something actually important: is a hot dog a sandwich? Yeah, I know, riveting stuff. But stick with me. I used this classic, pointless debate to test drive Google's Agent Development Kit (ADK), and it turned out to be a pretty decent way to show off what you can do with autonomous AI agents without writing a boatload of control logic.

The Setup: Herding AI Cats

The goal was simple: get AI agents to debate the hot dog question without me holding their hands. Using ADK, I set up a basic scenario:

The Mission: Settle the Hot Dog Question (or at least try).
The Squad:
- pro_debater: Argues for sandwich status. Its life depends on it. (Not really).
- con_debater: Argues against. Equally passionate. (Probably).
- moderator: Tries to keep the peace and summarize what just happened.
Rules of Engagement: Two rounds. Opening shots, rebuttals. Standard stuff.

Each agent got a system prompt – basically their marching orders. Told the pro_debater to make its case logically, gave the con_debater the opposite job. Simple.

(If you want the nuts and bolts of defining agent roles, the ADK Quickstart Guide is the place to start.)

How It Went Down

Predictably, it started with the basics.

Pro-Agent: "Look, definition says 'fillings between bread'. Bun's connected? So's a sub roll sometimes. It works."
Con-Agent: "Nah, the bun's one piece, designed to hold it. It's not two slices. Tacos have fillings and bread, you calling that a sandwich?"

They went back and forth like this for a bit. Pulled out dictionary definitions ("split roll" was a key phrase), argued about hinges vs. single units, cultural norms vs. strict definitions. The moderator agent actually did a decent job recapping who said what without screwing it up.

(Seeing how the ADK handles multiple agents talking and remembering the context is key. The ADK Samples Repository has examples showing different ways to set up these interactions.)

In the end, the pro-sandwich agent tapped out. Conceded that the "culinary context" and the bun's specific design made the con-agent's points solid enough. So, score one for Team Not-A-Sandwich in that run.

Why Bother? (The Actual Point)

Okay, the hot dog thing is just a hook. The interesting part is how this worked using ADK:

Set & Forget: Once I defined the agents and the flow, they ran the debate themselves. Less babysitting.
Defined Roles: The agents stuck to their assigned stances and argued accordingly.
Memory: They actually responded to each other's points from previous turns.
Process Control: The moderator kept the debate structure on track.

It's a trivial example, sure. But think about applying this framework to stuff that actually matters. Simulating customer service flows with different agent types. Generating code reviews from different 'expert' agents. Automating research summaries. Building game NPCs that don't just spout the same three lines. You get the idea. It's about orchestrating multiple AI instances to tackle a bigger task.

(Dive deeper into the ADK Samples GitHub Repository for more complex examples, or check the ADK API Reference if you want the gory details.)

Your Turn: Build Something (Maybe Win Stuff)

If this sounds interesting, or if you just like building things, Google's running an ADK Hackathon. Starts May 12th. There's a $50k prize pool, which isn't exactly chump change.

It's a good excuse to mess around with the ADK, build some multi-agent apps, and see what breaks. Whether you're experienced or just starting out with AI agents, it's worth checking out.

Sign up and get the details here: https://googlecloudmultiagents.devpost.com/

Go build something cool. Or, you know, finally solve the pineapple-on-pizza thing. Let me know how it goes.