◆ Autonomous Orchestrator
Pulls every thread on an escalated case … transaction flows, the entity network, adverse media, prior alerts … and assembles a coherent narrative of what happened and why it's suspicious. Spawns parallel sub-agents that fan out across every thread and merge into one link-analysis.
Memory
Working The case hypothesis and evidence board.
Episodic Linked prior cases and the entity's history.
Semantic Money-laundering typologies and red-flag indicators.
Store Knowledge graph (entity/transaction network) + case store
Orchestration
orchestrator-worker MCPA2A
Harness · Managed Agents … orchestrator spawning parallel research sub-agents (transaction analysis, network analysis, media); fresh context per sub-agent.
Tools
›_ Transaction graph queries Code exec ›_ Entity resolution / network analysis Code exec ⌘ Adverse-media + corporate registries MCP { } Case management API
Evals & guardrails
- Agent-as-judge review of narrative completeness vs. the evidence.
- Citation discipline … every claim in the packet links to a source record.
- An oversight agent approves the packet before anything proceeds to filing.
Offline reflection
Replays cases later confirmed (or cleared) to refine which red-flag combinations predict suspicious activity … sharpening the model, not just storing cases.
Frontier edge
- ▲Causal/counterfactual link analysis: tests whether the layering chain holds up under 'what if this hop were innocent?', so the narrative rests on a defensible chain of cause, not a suggestive coincidence of timing.
- ▲Long-horizon orchestration: runs a half-day, checkpointed investigation, fanning parallel research sub-agents across transaction, network and media threads and merging them into one cited timeline.
- ▲Self-improving fleet: red-flag combinations validated on confirmed cases are shared across the investigator population between runs (eval-gated), so the whole desk gets sharper, not just one case file.
A sample run
Trigger Escalation from the triage agent: confirmed structuring across three accounts.
- 1Spawn parallel sub-agents: transaction flow, entity network, adverse media.
- 2Merge findings into a timeline; identify the layering pattern.
- 3Quantify the suspicious sum and the deviation from expected activity.
- 4Draft the investigation summary with cited evidence.
Output A filing-ready packet handed to the SAR drafting agent with a recommended decision.
In numbers
1.2 hours
Median investigation latency
3,400
Cases closed / day