The Agentic Bank

Trade Surveillance Triage Agent

⬡ Tape Front-line disposition of trade-surveillance alerts at scale.
◆ Supervised Router

Reads each market-abuse alert, reconstructs the order-and-execution sequence, pulls the trader mandate and the prevailing market, and dispositions clear noise with a cited rationale. Scores against the spoofing, layering, wash-trade and front-running typologies, escalates anything above a conduct-risk threshold to the surveillance oversight agent, and retains a full reconstruction trace.

Memory

Working Per-alert scratchpad: the order sequence, hypotheses, disposition lean.
Episodic Prior alerts on the same trader/desk/instrument.
Semantic Market-abuse typologies (MAR/MAD) and the firm's surveillance scenarios.
Procedural Disposition playbooks refined from oversight-agent overrides.
Store Vector + time-series market context store

Orchestration

router-fanout MCPA2A

Harness · Managed Agents … session event-log per alert; context editing clears stale market-data dumps on long reconstructions.

Tools

{ } Surveillance platform API { } Order + execution records API { } Market data feed API Trader mandate / restricted lists Retrieval Escalation to surveillance oversight agent A2A

Evals & guardrails

  • Gold-set precision/recall vs. the labelled disposition set; agent-as-judge sampling.
  • Hard guardrail: cannot auto-close above a conduct-risk threshold … forced escalation to the oversight agent.
  • Full OpenTelemetry trace of every reconstruction retained for audit.
  • Monthly typology-drift detection against newly confirmed conduct cases.

Offline reflection

Nightly consolidation: replays oversight-agent overrides as Reflexion-style lessons, updating the disposition playbook (procedural memory).

Frontier edge

  • On-device edge inference: a distilled scorer runs on the surveillance feed itself, dispositioning clear noise in milliseconds before it ever queues.
  • World-model order-book simulation: replays the cancelled volume against a counterfactual book to test whether it could ever have executed … intent indicia, not just pattern match.
  • Eval-gated continual learning (SEAL-style): the disposition policy self-edits from each oversight-agent override and ships only after clearing the gold-set, so new manipulation typologies are absorbed without a full retrain.

A sample run

Trigger Spoofing alert: large sell orders placed and cancelled around a small buy fill.
  1. 1Reconstruct the full order lifecycle with timestamps against the order book.
  2. 2Check whether cancelled volume ever rested at executable prices.
  3. 3Pull the trader's mandate and recent alert history on the instrument.
  4. 4Score against the spoofing typology; draft rationale.
Output Escalates to the surveillance investigation desk with a pre-built packet … intent indicia and order-book impact quantified; clear-noise alerts auto-close with notes.

In numbers

9,800
Alerts dispositioned / day
91%
False-positive auto-clear rate
44s
Avg. time per alert

Handoffs

More on the Trade & Communications Surveillance desk