When Should Inference Be Split? A Fixed-Budget Theory of Predictable Multi-Agent Advantage under Local Context Ceilings

Abstract

This preprint develops a fixed-budget theory for when inference should be split across multiple agents under local context ceilings, yielding conditions for predictable multi-agent advantage over matched strong single-workspace baselines. It formalizes additive budget accounting across worker inference, routing, communication, memory, and verification, and derives diagnostics for candidate coverage, evaluation-selection accuracy, hijack risk, decomposability, diversity, shared-failure dependence, and communication fidelity.

Keywords

• fixed-budget inference
• multi-agent advantage
• local context ceilings
• test-time compute allocation
• matched single-agent baseline
• candidate coverage
• selection accuracy
• hijack risk
• communication fidelity
• external memory
• collective inference
• AI reasoning

Identifiers and source records

• DOI
• Works entry
• ORCID