Agent orchestration

The agent layer is not a single open-ended autonomous loop. Every workflow is an explicit LangGraph graph: stateful, long-running, conditionally routed, with human-in-the-loop checkpoints, retries and persistent execution state.

Why LangGraph

• Stateful, long-running workflows with persistent execution state.
• Conditional routing — the graph decides which agents run, in what order.
• Human-in-the-loop checkpoints where governance is required.
• Retries and deterministic, inspectable workflow structure.
• Multi-agent coordination instead of one giant prompt.

A research graph

Opportunity detected
        │
        ▼
Data completeness check ──insufficient──► collect more data
        │ sufficient
        ▼
Quant analysis → Security analysis → Social & narrative analysis
        │
        ▼
Memory retrieval → Multi-agent debate → Risk validation
        │
        ├── Rejected
        ├── Watch
        ├── Human review
        └── Approved

The scout surfaces candidates; the quant, security, on-chain and social agents enrich them; the debate agent tries to invalidate the trade; the risk agent can block it. Only what survives is scored and surfaced for governance.