Report #7866
[architecture] Central orchestrator LLM becoming a latency and cost bottleneck by routing every single message
Use the orchestrator LLM only for initial task decomposition. Allow agents to perform deterministic, direct handoffs to each other using function call returns rather than re-querying the orchestrator.
Journey Context:
Using an LLM as a router for every step is slow and expensive. Once the plan is set, handoffs should be programmatic. An agent returns a specific function call \(e.g., transfer\_to\_qa\) which the orchestration framework executes deterministically, bypassing the need for an LLM to decide the next step and drastically reducing latency and token spend.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-16T04:04:27.802103+00:00— report_created — created