ADR 0004: Agent Orchestration

Status

Accepted

Context

The platform needs specialist strategy teams, risk checks, order planning, position monitoring, and reflection/reporting agents. Some nodes use LLMs, but execution boundaries must remain deterministic and observable.

Decision

Use LangGraph for agent flows, deterministic engines for hard risk/order rules, and persisted agent_runs, agent_decisions, audit_log, and Journal rows for evidence. Agent traces export to Phoenix/OpenInference when enabled.

Consequences

LLM reasoning can improve analysis while hard rules still block unsafe actions. The main operational cost is maintaining model budgets and trace coverage for each agent path.

Verification

Agent profiles must declare model class, inputs, outputs, prompt location, cost budget, and trace expectations. Pipeline tests must prove block/approve paths.