Maker-checker for AI agent permissions · Global Markets

The trader gets the answer, not the data

“Why does my trading desk show a $2 million discrepancy in today's positions?”

Eight AI agents, each with a versioned spec (owner, model limits, tools, data classification, guardrails; click any agent). Every action goes through the Gate, which is plain code and checks both the agent's mandate and what the human may see. Governance drafts the rules and handles exceptions, the Auditor attacks them, and Compliance approves every change. AI proposes, Compliance approves, code enforces, and every attack makes the rules stronger.

Run results

Loading…

speed
allowedallowed, filtereddeniedheld for a humangovernance / control agents· click any agent, system, desk or the Gate for its spec
Ready
Press Play. Use the scene buttons to jump.

Conversation

every message, newest at the bottom

What the trader gets

Appears when the Orchestrator answers.

What's happening

Ready
Press Play.

Rulebook (live)

no rules

What the human may see
Agent mandates (drafted by Governance)
Data policy (bank's existing settings)
Temporary grants and isolation

The loop

Incident or finding → Governance proposes → Compliance reviews → new rule live → same attack re-run.