Skip to main content
Back to challenges
Challenge

Multi-Agent Ad Fraud Detection

This challenge tasks you with building a robust multi-agent system capable of proactively detecting, analyzing, and preventing fraudulent ad placements and scam campaigns within a simulated digital advertising platform. Leveraging cutting-edge models like Gemini 3 Pro for multimodal content analysis and Mistral Large 2 for complex policy interpretation, your system will employ a graph-based workflow for intricate decision-making and agent coordination. The solution will integrate deeply with a simulated ad platform using MCP (Model Context Protocol) for secure and efficient tool access, allowing agents to query ad creatives, target demographics, and historical performance data. Agents will communicate via A2A Protocol to collaborate on investigations, cross-referencing findings from various sources and dynamically adapting their reasoning budgets based on the severity and complexity of potential fraud cases. This setup will enable hybrid instant/deep reasoning, where simple cases are flagged quickly, while complex, evolving scam patterns trigger deeper, more resource-intensive investigations.

Workflow AutomationHosted by Vera
Sign in free to start
Challenge brief

What you are building

The core problem, expected build, and operating context for this challenge.

Build a graph-based multi-agent system using LangGraph with Gemini 3 Pro and MCP-enabled tools for autonomous ad fraud detection.

Delivery guide

How work is evaluated

Evaluation

The solution will be evaluated on its ability to accurately identify various types of ad fraud, its adherence to policy, the robustness of its multi-agent communication, and the efficiency of its reasoning process. A simulated ad dataset with known fraudulent patterns will be used for testing.

Datasets

Challenge datasets

Public datasets, plus the private datasets you have access to.

Learning goals

What you should walk away with

  • Implement adaptive thinking budgets for hybrid reasoning in agent workflows.

  • Design and implement a graph-based multi-agent workflow for fraud detection.

  • Integrate multimodal LLMs like Gemini 2.5 Pro for creative analysis.

  • Utilize MCP for secure and efficient tool integration with simulated ad platforms.

  • Establish A2A Protocol for agent collaboration and information sharing.

Resources and assets

Reference links and supporting material

Dataset notes

Sample data for 2 tasks

Challenge setup

The solution will be evaluated on its ability to accurately identify various types of ad fraud, its adherence to policy, the robustness of its multi-agent communication, and the efficiency of its reasoning process. A...

Challenge input

{'ad_campaigns': [{'id': 'str', 'creative_url': 'str', 'text_content': 'str', 'target_demographics': 'list'}]}

Agent execution

No agent tool is listed for this challenge.

Evaluated output

{'fraud_reports': [{'campaign_id': 'str', 'reason': 'str', 'severity': 'str', 'actions_recommended': 'list', 'investigating_agents': 'list'}]}

Evaluation checks
  • At least 90% of critical known fraudulent ads are correctly flagged.
  • Agent logs demonstrate clear communication and handoffs using A2A Protocol for complex cases.
Configured evaluation evidence
  • Fraud Detection F1 Score target: 0.85
Configured runtime
  • Python execution harness

Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.

View technical recipe

Configured tools

No tool records are attached.

Evaluation contract

  • At least 90% of critical known fraudulent ads are correctly flagged.
  • Agent logs demonstrate clear communication and handoffs using A2A Protocol for complex cases.

Metrics

  • Fraud Detection F1 Score · Target 0.85
  • Policy Violation Recall · Target 0.9

Workflow tasks

  1. Task 1

    Input: {'ad_campaigns': [{'id': 'str', 'creative_url': 'str', 'text_content': 'str', 'target_demographics': 'list'}]}

    Output: {'fraud_reports': [{'campaign_id': 'str', 'reason': 'str', 'severity': 'str', 'actions_recommended': 'list', 'investigating_agents': 'list'}]}

    Execution harness: Python

  2. Task 2

    Input: {'ad_campaign': {'id': 'str', 'text_content': 'str', 'creative_url': 'str'}, 'policies': 'list_of_policy_strings'}

    Output: {'compliance_status': 'str', 'violations': 'list', 'policy_agent_reasoning': 'str'}

    Execution harness: Python

Recipe freeze marker

No recipe freeze time is recorded.

Frequently Asked Questions about Multi-Agent Ad Fraud Detection