◆ Autonomous Router
Reads the firm's email, chat and transcribed voice for conduct and information risk: off-channel dealing, MNPI leakage, collusion, harassment. Resolves lexicon false positives in context (a 'kill the trade' that is just a cancelled order), reasoning over the full thread and participants. Routes concern flags to the conduct investigation agent; never auto-dispositions a flag alone.
Memory
Working The message + its conversation thread and participants.
Episodic Prior flags on the same participants or trading book.
Semantic Conduct typologies, information-barrier rules, firm slang and aliases.
Store Vector store + participant graph
Orchestration
router-fanout MCP
Harness · Managed Agents … high-throughput batch sessions; per-thread context assembly.
Tools
{ } Comms archive (e-comms + voice) API { } Transcription service API ⌕ HR + trading-book directory Retrieval ⇄ Escalation to conduct investigation agent A2A
Evals & guardrails
- Gold-set precision/recall on a labelled conduct corpus; agent-as-judge sampling.
- Privacy guardrail: review scope and access logged immutably; no scope creep.
- Concern flags route to the conduct investigation agent … never auto-dispositioned alone.
Frontier edge
- ▲Confidential compute: reads e-comms and transcribed voice inside a secure enclave, so employee PII is never in clear and the access boundary is hardware-enforced.
- ▲Voice-native multimodal: scores tone, hesitation and turn-taking in audio directly rather than relying on a lossy transcript, catching the coded language a text pass would flatten.
- ▲Continual learning on firm slang: eval-gated self-edits keep pace with new aliases and desk in-jokes without a labelling cycle, so the lexicon never goes stale.
Handoffs
Hands to → Conduct Investigation Agent