AI Agent Rules — BANXE AI BANK

Rule ID: 80-ai-agents | Load order: 80

Created: 2026-04-12 (IL-RETRO-02) | IL-ARL-01, IL-SK-01, IL-069..071

Agent Types

Soul File Structure

Every new AI skill requires a .soul.md file at agents/compliance/soul/<agent-name>.soul.md:

# <AgentName> Soul — BANXE AI BANK
## Identity
## Capabilities
## Constraints (MUST NOT / MUST NEVER)
## Autonomy Level (L1-L4, see agent-authority.md)
## HITL Gates (which decisions require human approval)
## Protocol DI Ports (which external ports this agent uses)
## Audit (what it logs and where)

Orchestrator Registration

New agents must be registered in agents/compliance/orchestrator.py:

AGENT_REGISTRY = {
    "my-agent": MyAgent(port=InMemoryMyPort()),
    ...
}

Protocol DI for Agents

Agents use the same Protocol DI pattern as services:

class MyAgentPort(Protocol):
    async def fetch_data(self, id: str) -> dict: ...

class InMemoryMyPort:
    async def fetch_data(self, id: str) -> dict:
        return {"id": id, "stub": True}

class MyAgent:
    def __init__(self, port: MyAgentPort) -> None:
        self._port = port

    async def run(self, context: dict) -> AgentResult:
        data = await self._port.fetch_data(context["id"])
        ...

HITL Gates in Agent Code

Agents NEVER auto-apply changes above their autonomy level. Pattern:

if self.autonomy_level >= AutonLevel.L3:
    # auto-process
    return self._process(data)
else:
    # HITL: propose, don't apply
    return HITLProposal(
        action=data,
        requires_approval_from="MLRO",
        reason="exceeds L2 autonomy threshold",
    )

See services/hitl/hitl_service.py for the gate implementation.

Swarm Pattern

Multi-agent workflows use agents/compliance/swarm.yaml:

• Defines agent roles, trust levels, escalation paths
• All swarm decisions go through SwarmOrchestrator
• No agent calls another agent directly — only via orchestrator

Agent Autonomy Levels (from agent-authority.md)

Always set the lowest acceptable autonomy level. Err towards L2 over L1.

ARL (Agent Routing Layer)

All multi-step agent tasks route through route_agent_task MCP tool:

• Tier 1: Claude Haiku (fast, simple tasks, routing/classification)
• Tier 2: Claude Sonnet (standard tasks, analysis)
• Tier 3: Claude Opus (complex tasks, financial decisions)

Routing decision is logged to ClickHouse (I-24). Never bypass ARL for production agent calls.

Compliance Knowledge Base Integration

Agents accessing regulatory guidance must use the KB protocol:

class KBQueryPort(Protocol):
    async def query(self, question: str, jurisdiction: str = "UK") -> list[Citation]: ...

Direct ChromaDB access from agents is forbidden — use KBQueryPort or MCP tool kb_query.

Invariants for Agent Code

References

• Swarm config: agents/compliance/swarm.yaml
• Soul directory: agents/compliance/soul/
• HITL service: services/hitl/hitl_service.py
• Agent registry: .ai/registries/agent-map.md
• ARL entry: services/arl/ (commit 5f132dd)
• Agent authority: .claude/rules/agent-authority.md
• ADR: docs/adr/ADR-004-fastmcp-agent-tooling.md, docs/adr/ADR-005-protocol-di-pattern.md