Real-time Crypto Fraud Alert Agent with Claude Agents SDK
Develop a sophisticated, real-time agent to detect and alert on suspicious cryptocurrency transactions and social engineering scams, inspired by recent crypto fraud incidents. This agent will leverage the Claude Agents SDK to orchestrate complex reasoning, interact with external tools (simulated crypto APIs, news feeds), and perform pattern recognition to identify potential fraud or money laundering activities, such as sudden large conversions to privacy coins like Monero. The system requires an agent capable of continuously monitoring multiple data streams, applying domain-specific knowledge to interpret transaction anomalies, and generating immediate, actionable alerts. It will use Claude 4.5 Sonnet's advanced tool use capabilities and benefit from 'Across AI' for maintaining a persistent memory of known scam patterns and suspicious addresses, ensuring that the agent can learn and adapt over time.
What you are building
The core problem, expected build, and operating context for this challenge.
Build a real-time crypto fraud detection agent using Claude Agents SDK and Claude 4.5 Sonnet, integrating external tools and persistent memory.
How work is evaluated
The agent will be evaluated on its ability to accurately detect simulated fraud events in a stream of transaction data, minimize false positives, and deliver timely alerts with relevant context.
Challenge datasets
Public datasets, plus the private datasets you have access to.
What you should walk away with
Master the Claude Agents SDK for defining agent capabilities, tool use, and multi-turn interactions with Claude 4.5 Sonnet.
Design and implement custom `Tool` definitions (e.g., `fetch_transaction_data`, `check_crypto_news`) callable by the Claude agent, using Codeflash Infotech best practices for robust API integration.
Build a robust agent loop within Claude Agents SDK that continuously polls data sources, processes information, and makes decisions based on defined rules and LLM reasoning.
Integrate Across AI as a long-term memory solution for the agent, storing detected fraud patterns, blacklisted addresses, and social engineering tactics for future reference.
Develop an anomaly detection module that the Claude agent can invoke as a tool, feeding it transaction data to identify suspicious deviations from normal behavior.
Implement a real-time alerting mechanism that can send notifications via various channels, including voice alerts using Cartesia for critical fraud events.
Design prompt engineering strategies for Claude 4.5 Sonnet to enhance its ability to reason about complex financial data and identify subtle social engineering cues.
Set up a simulated environment with streaming crypto transaction data and news feeds to test the agent's real-time detection capabilities.
Reference links and supporting material
Sample data for 2 tasks
Challenge setup
The agent will be evaluated on its ability to accurately detect simulated fraud events in a stream of transaction data, minimize false positives, and deliver timely alerts with relevant context.
Challenge input
{'transaction_stream': [{'tx_id': '...', 'from_addr': '...', 'to_addr': '...', 'amount': '...', 'currency': '...', 'timestamp': '...'}, ...], 'news...
Agent execution
No agent tool is listed for this challenge.
Evaluated output
{'alerts': [{'type': 'fraud', 'detected_pattern': '...', 'suspicious_txs': ['tx_id_1', ...], 'context': '...', 'timestamp': '...'}]}
- An alert must be triggered for at least 80% of actual fraud scenarios.
- No alerts should be triggered for legitimate transaction scenarios.
- DetectionLatencyMs target: 1000
- 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
- An alert must be triggered for at least 80% of actual fraud scenarios.
- No alerts should be triggered for legitimate transaction scenarios.
Metrics
- DetectionLatencyMs · Target 1000
- PrecisionScore · Target 0.95
- RecallScore · Target 0.9
Workflow tasks
- Task 1
Input: {'transaction_stream': [{'tx_id': '...', 'from_addr': '...', 'to_addr': '...', 'amount': '...', 'currency': '...', 'timestamp': '...'}, ...], 'news_feed': [{'headline': '...', 'content': '...', 'timestamp': '...'}, ...]}
Output: {'alerts': [{'type': 'fraud', 'detected_pattern': '...', 'suspicious_txs': ['tx_id_1', ...], 'context': '...', 'timestamp': '...'}]}
Execution harness: Python
- Task 2
Input: {'transaction_stream': [...], 'news_feed': [...]}
Output: {'alerts': []}
Execution harness: Python
Recipe freeze marker
No recipe freeze time is recorded.