Switchboard: Privacy-Aware Hybrid Routing Across Heterogeneous AI Agents

Switchboard is a privacy-aware, local-first orchestration layer that routes eachrequest across a heterogeneous mix of AI agents — subscription-backed CLI codingagents (Codex, Claude Code), reasoning assistants, local open-weight models, anddeterministic tools. It targets the constraints prior cost-aware routing ignores —scarce subscription quota, hard privacy requirements, and context fragmented acrossincompatible agent interfaces — using a deterministic policy that always overridesthree small local classifiers, plus structure-aware context compression, localsemantic memory, and metadata-only telemetry. On a 100-case benchmark across multiplereal-hardware runs, the learned hybrid router preserves most premium-agent answerquality at about 38% premium usage with no observed private-prompt leak. This deposit (version 2) adds a reproduction bundle alongside the manuscript: themulti-run experiment and statistical-aggregation harness, the per-case records of allruns, and the aggregated tables and figures. Open-source:https://github.com/aivinay/switchboard (PyPI: switchboard-local)This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible.