Parallel Multi-Persona Deliberation with a Measured-Token Circuit Breaker and Degrade-to-Best-Single
Abstract
This document discloses, as public prior art, an orchestration mechanism that consults the top-N ranked personas (or agents) of a multi-agent AI system concurrently, and then decides whether to execute a costly downstream synthesis stage on the basis of the actual number of prompt and completion tokens the parallel round has already consumed — a quantity measured after the calls return, not estimated before they are issued. Each per-persona consultation runs under an independent Promise.race deadline; a failed or timed-out consultation resolves to a structured, zero-token record rather than throwing, so the orchestrator can always compute a complete per-persona ledger. Once the round completes, the orchestrator sums the measured tokens and compares them against a configurable budget. If the measured spend exceeds the budget, a circuit breaker trips: the orchestrator skips synthesis entirely and returns the best already-paid-for single answer, tagging the audit trace circuitBroken. Where the breaker does not trip, a layered degradation ladder governs the remaining cases — zero valid answers, exactly one valid answer, and synthesis failure — each producing a distinct, machine-readable trace state. The novelty over the mixture-of-agents (MoA), LLM-Blender, and multi-agent-conversation art is the mid-pipeline metering of consumed tokens used to deterministically cut the synthesis stage; the frameworks in the prior art budget at the request boundary or not at all. A clean-room, dependency-free, offline-runnable reference implementation accompanies this disclosure.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Assuncao, gustavo matthew, "Parallel Multi-Persona Deliberation with a Measured-Token Circuit Breaker and Degrade-to-Best-Single", Technical Disclosure Commons, (July 13, 2026)
https://www.tdcommons.org/dpubs_series/10876