Build a Cross-Border Sanctions Triage Agent with Claude Agents SDK and Devin
Lifting or re-imposing international state sanctions requires rapid reconfiguration of global payment screening engines. Develop a Claude Agents SDK agent using Cognition Devin workflows to evaluate cross-border transactions against evolving OFAC and UN regulatory directives. Achieve sub-100ms classification latency while maintaining zero false negatives for sanctioned entities.
What you are building
The core problem, expected build, and operating context for this challenge.
Construct an autonomous sanctions compliance triage system using Claude Agents SDK and Cognition Devin to parse dynamic policy updates and filter transaction streams.
How work is evaluated
Verify that updated state policy shifts dynamically change agent transaction approvals without manual system restarts.
Challenge datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
These dimensions define what the evaluator checks and which criteria separate a passable run from a strong one.
Zero Sanction Leakage
Ensures no transactions from active prohibited entity list are approved.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
False Negative Rate
Rate of non-compliant transactions accidentally approved. • target: 0 • range: 0-0
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
What you should walk away with
Implement sanctions rules dynamic reloading using Claude Agents SDK
Automate code patch updates in payment gatekeepers via Cognition Devin
Ensure real-time parsing of ISO 20022 payment messages against dynamic watchlists
Evaluate cross-border routing risks when sovereign entity classifications change
Reference links and supporting material
Synthetic benchmark of 2,500 ISO 20022 cross-border payment XML payloads targeting sanctioned and recently un-sanctioned jurisdictions.
Challenge setup
Verify that updated state policy shifts dynamically change agent transaction approvals without manual system restarts.
Challenge input
JSON containing sender_country, receiver_country, intermediary_banks, and payload_details.
Claude Agents SDK
Provides rigorous extended thinking for complex legal and regulatory decision trees.
Cognition Devin
Automates updating rules engines and screening microservices in code repositories.
Evaluated output
JSON with decision ('APPROVE', 'HOLD', 'REJECT'), policy_applied, and confidence.
- Ensures no transactions from active prohibited entity list are approved.
- Rate of non-compliant transactions accidentally approved. • target: 0 • range: 0-0
- Benchmark: SanctionsComplianceBench
- False Negative Rate target: 0%
- 1 public reference case
- Python execution harness
- Python sandbox (unavailable on Versalist)
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
- Claude Code · Required
- Cognition Devin · Optional
- Claude Code · Required
- Cognition Devin · Optional
- Claude 4.1 Opus · Optional
- Cognition Devin · Optional
Evaluation contract
- Zero Sanction Leakage · Weight 1
- False Negative Rate · Weight 1
Metrics
- False Negative Rate · Target 0
Benchmarks
- SanctionsComplianceBench
- security
Reference cases
- 1 public reference case
Workflow tasks
- Task 1
Input: JSON containing sender_country, receiver_country, intermediary_banks, and payload_details.
Output: JSON with decision ('APPROVE', 'HOLD', 'REJECT'), policy_applied, and confidence.
Execution harness: Python
Evaluation tools
- Claude Agents SDK · Agent Framework
- Cognition Devin · Autonomous Software Engineer
Recipe freeze marker
No recipe freeze time is recorded.