Multi-Agent Cost Tradeoffs & the "Not Automatically Better" Trap

Multi-agent designs are powerful but not free, and the exam tests whether you reserve them for work that actually needs them.

The Real Cost

The cost of a manager/subagent hierarchy is real: multiple agents multiply token usage and add coordination overhead. Every subagent invocation is its own set of model calls, and the manager must spend additional turns delegating, collecting, and synthesizing results. Reserve multi-agent hierarchies for work that is genuinely separable and expensive enough to justify the overhead -- not as a default.

Common Exam Trap

> Trap: Subagents are not "just more prompts." Their defining benefit is a separate context window (a first-class fix for context bloat). And a multi-agent design is not automatically better -- a single well-scoped augmented LLM often wins on cost and latency.

This trap has two halves that are easy to conflate on scenario questions:

1. Don't undervalue subagents by treating them as just extra prompts -- their real value is context isolation.
2. Don't overvalue multi-agent designs by assuming more agents means better quality -- a single augmented LLM is frequently the cheaper, faster, equally correct answer.

Decision Checklist

Before recommending a multi-agent design, confirm:

• Is the task genuinely separable into subtasks that benefit from isolated context, specialization, or parallel execution?
• Is the task expensive/heavy enough (e.g., broad research spanning many sources) that the added token and coordination cost is justified?
• Would a single augmented LLM call, or a simpler workflow, satisfy the same requirement at lower cost and latency?

If the answer to the third question is yes, the multi-agent answer is a trap, regardless of how impressive it sounds.