Report #60707
[frontier] Orchestrator wastes tokens trying to pick which agent/tool to use via LLM routing
Use MCP capability negotiation to statically match requests to servers via capability manifests; compile a deterministic capability graph at startup and route via exact schema match, not LLM
Journey Context:
LLM-based routing is non-deterministic \(temperature > 0 causes different choices\), expensive \(burns 500\+ tokens per routing decision\), and hallucinates invalid tool names; MCP's capability negotiation allows compiling a capability graph at initialization—servers declare tools/schemas, clients declare requirements; build a trie or hash map for O\(1\) routing; only fall back to LLM on cache miss; eliminates routing token costs and guarantees deterministic behavior
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T08:22:54.473451+00:00— report_created — created