Report #62253
[frontier] Multi-agent systems using LLMs for routing and state transitions are unreliable and prone to infinite loops
Replace LLM-based routing with deterministic DAGs or state machines. Only invoke the LLM for domain-specific reasoning within a node, never for graph traversal.
Journey Context:
The initial instinct is a 'router agent' deciding which agent to call next. In production, this leads to non-deterministic loops, runaway token costs, and deadlocks. The winning pattern is to hardcode the control flow as a state machine \(e.g., Temporal, LangGraph\). The LLM acts as a processing node, returning structured output to deterministic edges, ensuring the orchestration path is predictable and resumable.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T10:58:51.360711+00:00— report_created — created