Module IV — Single-Agent vs Multi-Agent

Status: outline. Lecture body not authored.

Central question:

When do we actually need another agent?

Recent literature shows that, under equal reasoning-token budgets, a single-agent system can outperform a multi-agent system on multi-hop reasoning tasks. That does not prove universal single-agent superiority, but highlights coordination overhead.

Architectures

Supervisor / Workers · Planner / Executor · Planner / Implementer / Reviewer · Generator / Evaluator · Maker / Checker · Hierarchical · Peer-to-Peer · DAG · Blackboard · Swarm · Competitive ensemble.

Agent Boundary Cost

Every additional agent is not an incremental capability gain by default; it establishes an explicit operational boundary:

MORE AGENTS ≠ MORE CAPABILITY BY DEFAULT

ADDITIONAL AGENT
        ↓
NEW CONTEXT BOUNDARY
        ↓
NEW COORDINATION COST
        ↓
NEW FAILURE SURFACE

Each agent boundary incurs five concrete costs:

1. Context Boundary: An isolated token window requiring selective assembly, compaction, and context assembly/injection. 2. Serialization / Deserialization: Overhead and schema risks during data conversion between agents. 3. Semantic Drift: Cumulative mutation of intent and constraints across intermediate representations. 4. Authority Boundary: Loss of epistemological distinction between verified facts, agent deductions, and memory. 5. Coordination Overhead: Latency, token expenditure, and synchronization bottlenecks across workers.

Unique process — Agent Necessity Test

Before adding an agent, a structural justification must exist:

SPECIALIZATION
PARALLELISM
ISOLATION
VERIFICATION
FAULT TOLERANCE
CAPABILITY ACCESS

Evaluation Rule: An agent is justified only if the architectural benefit along one of these axes materially exceeds the Agent Boundary Cost (coordination overhead and transfer failure surface).

If none apply, or if boundary risks dominate: keep Single-Agent.