The Second Agent Test: Why Multi-Agent ROI Begins by Proving That One Agent Was Not Enough

Large-language-model agent systems are increasingly designed as teams: planners, researchers, critics, coders, verifiers, supervisors, routers, and specialists exchange messages and call tools in pursuit of a shared task. The dominant engineering question is how to coordinate these agents. This paper argues that an earlier question must be answered first: why should a second agent exist at all? A second capability does not necessarily justify a second agent. The same capability may be represented as an instruction, deterministic function, tool, skill, internal role, workflow node, or agent-as-tool without creating another governed center of state, authority, evidence, failure, accountability, and return. The central thesis is that multi-agent architecture is an earned status. A separate agent is qualified only when the work requires a boundary that no lower separation level can preserve adequately, and when the accepted value or expected loss avoided by that boundary exceeds its full creation, coordination, governance, and recovery cost. The paper introduces the Dynamic Multiplicity Qualification Architecture, the Separation Ladder, the Agenthood Qualification Contract, the Multiplicity Qualification Contract, the Agenthood Lease, the Boundary Loss Counterfactual, the Agent Boundary Balance Sheet, and two connected operators for design-time and runtime qualification. It defines provisional laws for capability-agent separation, lowest sufficient separation, counterfactual baselines, non-compressibility, best-member preservation, effective agent count, boundary value, lease expiration, promotion, demotion, and multiplicity elasticity. Three benchmark constitutions are proposed: the Second Agent Qualification Benchmark, the Agenthood Qualification Benchmark, and the Agent Boundary Value Benchmark. The paper also states theorem schemas, negative theorem schemas, falsifiers, benchmark constitutions, and non-claims. Current research reports both benefits and failures of multi-agent systems: improved reasoning, heterogeneous evidence, parallelism, specialization, and adaptive topology in some settings, but coordination failures, correlated errors, compute confounding, homogeneous saturation, communication burden, and single-agent dominance under matched budgets in others. These mixed findings support qualification rather than doctrine. The contribution is conceptual, economic, and architectural. It does not claim that single agents are universally superior, that multi-agent systems are merely hype, that every visible agent represents an effective boundary, or that authority, legitimacy, evidence independence, recovery, and optionality can be reduced to one universal financial formula. Instead, it proposes a bounded research and enterprise decision architecture for determining when another agent is necessary, what boundary it preserves, how that boundary changes accepted flow value, loss exposure, exercisable option value, and full incremental cost, which non-compensatory gates constrain its promotion, how it interacts with the wider boundary portfolio, how long it deserves to remain separate, and whether its continued existence returns more intelligence than it consumes.