Agent Routing — References

There is no settled literature on selecting between agents. The closest established work is on gating between specialised components and on selecting an executor from a large catalogue.

Gating and expert selection

• Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer (Shazeer et al., 2017), arXiv:1701.06538 — a gating function choosing among specialists, including the load-imbalance problem where a few experts attract most of the traffic. The failure is structurally the same one a catalogue of agents exhibits.

Choosing an executor from a catalogue

• Gorilla: Large Language Model Connected with Massive APIs (Patil et al., 2023), arXiv:2305.15334 — selecting the correct callable from a large, changing catalogue, and the measurement of hallucinated selections as a distinct error class.
• ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs (Qin et al.) — selection and invocation at catalogue sizes where putting every description in the prompt stops being possible. Cited by name.

Abstention

• Semantic Router — an embedding-similarity dispatch library with explicit score thresholds, useful mainly as a concrete illustration of why a confidence floor and a no-match path are required. Cited by name.

No verified reference is recorded for evaluating agent selection quality in production; the practice in use is shadow comparison and manual adjudication, described in the pattern note rather than drawn from a published source.