Scamper

Live Demo: https://scamper-e1d5d599d982.herokuapp.com/

username: judge

password: ethglobal

Scamper is an AI agent whose goal is to identify scam tokens and protect crypto investors. It’s built with a framework that closely integrates Coinbase AgentKit and the OpenAI Agents SDK][openai-agents-sdk, providing out-of-the-box web search, thread history, and a web API + frontend interface.

In addition, Scamper employs a multi-agent research loop when asked about a token. Here’s how it works:

1. Scamper (the main agent) receives the user prompt (e.g., “Is XYZ token a scam?”).
2. Coordinator Agent splits the request into subtopics and shares them with specialized Researchers.
3. Researchers perform web searches, gather data, and summarize.
4. Coordinator combines their reports into a concise final answer.
5. The final response is handed back to Scamper, who presents it to the user.

This mini‐ecosystem is a powerful demonstration of the OpenAI Agents SDK for building sophisticated multi-agent solutions with minimal boilerplate. Scamper's integration with Coinbase AgentKit means these sophisticated multi-agent flows can be used to control onchain agentic applications.

Features

• Scam Detection: Flag suspicious tokens or tokens with red flags.
• Web-Search Enabled: Researchers leverage integrated search to fetch live info.
• Thread History: The conversation context is maintained for a more natural flow.
• Multi-Agent Orchestration: Dynamic creation of sub-agents (Coordinator and multiple Researchers).

How It Works User --> Scamper (Main Agent) --> multi_agent_research (Tool) | +--> Coordinator Agent --> (1) Subtopic "X" | -> (2) Subtopic "Y" +--> Researcher Agents perform searches, short reports | +--> Final summary to user

Scamper is the user-facing persona, deciding whether more in-depth research is needed. Coordinator breaks down the question into subtopics. Researchers each do a web search on their assigned subtopic, returning concise findings. Coordinator merges the findings into one final summary, which Scamper then sends to the user.

We combined AgentKit and OpenAI's Agents SDK to create multi-agent workflows, and made an upstream code contributions to AgentKit to enable web search functionality by default when using the Agents SDK. We've also submitted our Docker image to Autonome, and hope to make our API interface compliant with their standard so we can be available in the Autonome store.

Our greatest hope coming out of this is that we make it much easier for other developers to create advanced web-based agents on top of AgentKit and the Agents SDK! Our framework combines the two into a simple web-based system, including a frontend chat interface and a prompt API. Now that the framework is built, it will be easy to quickly iterate on new agentic experiments.