Why and When to Use Multiple Agents

Specialisation has a price

A single agent with too many tools and a long prompt gets confused: it picks the wrong tool, forgets rules and becomes hard to test. Splitting the work into specialised agents (a researcher, a coder, a reviewer, a billing expert), each with a short prompt and a small tool set, can improve reliability, let you use different models per role and allow independent testing and ownership. The costs are real: more model calls, more latency and tokens, coordination bugs (agents repeating work or contradicting each other), harder debugging and a larger attack surface. Start with one agent or a workflow graph; split only when a single agent measurably fails.

Specialists and a coordinator

Split a big job among focused agents, then coordinate them with a supervisor or hand-offs.

One agent versus many

A rough comparison to guide the decision.

                      single agent / workflow         multiple agents
prompt + tools        one big set (can confuse)       small, focused per role
model choice          one model                       cheaper/stronger per role
cost, latency         lowest                          higher (more calls, more tokens)
debugging             one trace                       many traces + hand-offs
failure modes         wrong tool, forgotten rules     loops between agents, duplicated work
start here?           YES                             only after measuring a single-agent failure

Quick check: What is a real cost of multi-agent systems?

• They are always cheaper
• More calls, latency and coordination bugs
• They remove the need for testing
• They need no prompts