the fraud platform AI-Native Fraud Investigation Platform — Agent Sequence
A flagged fraud case triggers an agentic workflow. Claude autonomously decides which tools to call,
gathers evidence in parallel, synthesizes findings, and produces a report with confidence levels.
Every tool call is traced to LangFuse. Sensitive actions require analyst approval.
sequenceDiagram
participant P as Fraud Platform
participant K as Kafka
participant A as Claude Agent
participant T1 as closest_user_search
participant T2 as cluster_search
participant T3 as transaction_explainer
participant T4 as model_explainer
participant L as LangFuse
participant AN as Analyst
P->>K: case.flagged
Note over P: $2,499 crypto purchase 2-hour-old account device overlap detected
K->>A: Start investigation workflow
activate A
Note over A: Evaluate case context Select tools to call
par Parallel tool calls
A->>T1: user_id=usr_f83a1c
activate T1
T1-->>A: 2 matches (91%, 84%)
deactivate T1
A->>L: span: closest_user_search
and
A->>T2: user_id=usr_f83a1c
activate T2
T2-->>A: ring_12 (3 accounts)
deactivate T2
A->>L: span: cluster_search
and
A->>T3: txn_id=TXN-98231
activate T3
T3-->>A: high-risk + velocity spike
deactivate T3
A->>L: span: transaction_explainer
and
A->>T4: case_id=CASE-4871
activate T4
T4-->>A: device 34%, age 27%
deactivate T4
A->>L: span: model_explainer
end
Note over A: Synthesize evidence Separate facts vs signals Generate recommendation
A->>L: recommendation: BLOCK (95%)
A->>AN: Investigation report
deactivate A
Note over AN: Review evidence Approve/reject action
AN->>K: case.action_confirmed
Note over AN: Block transaction Freeze ring members
Turn 1: Claude calls all 4 tools in parallel. Turn 2: Synthesize evidence into structured report with confidence level.
Observability
Full LangFuse trace
Root trace: investigation-CASE-4871 Child spans per tool call with input/output captured.
Guardrails
Evidence-grounded output
Every finding cites which tool produced it. Unsupported claims are blocked. Sensitive actions flagged for analyst approval.
Output Structure
Facts • Signals • Recommendation
Verified facts from tools. Model-derived risk signals. Recommendation with confidence level. Approval requirement.
Try It
fraud-platform crypto_new_account --verbose
Run the working prototype to see this exact workflow in action. Traces sent to LangFuse automatically.
This is not a hypothetical diagram. The sequence above matches the actual agent prototype output.
Claude calls all four the fraud platform tools, synthesizes evidence, and produces the investigation report shown here.
Run fraud-platform crypto_new_account --verbose to see it live.