LEAD-01-025-P · Primer — Single-Agent vs Multi-Agent

The idea in one sentence

A single agent handles one goal with one loop and one context; a multi-agent system splits the work across specialized agents that coordinate — and the extra power comes with extra ways to fail.

Why it exists

Multi-agent architectures are seductive because org charts are how we scale human work, so a team of specialist bots feels natural — a researcher, a writer, a checker. But every hand-off between agents is a place for context to be lost and errors to compound, exactly like a company where four departments each do their part and the customer's problem still falls through the cracks. The reliability↔autonomy tension is sharpest here: more agents means more capability and less predictability, so you add them only when one agent genuinely cannot hold the job.

How it works

Start single-agent by default: one goal, one context window, one loop, one place to debug. Reach for multi-agent only when you hit a real limit — the task needs genuinely different skills or tools, the context is too large for one agent to hold, or you want independent review (one agent does the work, a separate evaluator agent checks it, which mirrors separation of duties). When you do split, use a clear coordination pattern: a supervisor that delegates and integrates, or a fixed pipeline of specialists, not a free-for-all chat where agents talk in circles. Every hand-off needs a crisp contract — what goes across and how success is judged — because coordination cost, not model quality, is what usually kills these systems. Example: a due-diligence workflow may use a retrieval agent, an analysis agent, and a separate red-team agent that argues the opposite case, precisely because independent challenge is worth the extra complexity.

See it

flowchart TD
  A["New Task"] --> B{"One skill +<br/>fits one context?"}
  B -->|yes| C["Single Agent"]
  B -->|no| D["Supervisor"]
  D --> E["Specialist:<br/>Research"]
  D --> F["Specialist:<br/>Analyze"]
  D --> G["Reviewer<br/>Agent"]
  E --> D
  F --> D
  G --> D

In practice

Lead qualification: a single agent can enrich, score, and route most leads fine; you only split off a dedicated research agent when high-value enterprise leads need deep company investigation that would blow past one agent's context — and even then a supervisor stitches the findings back together.

The one thing to remember

Add a second agent for the same reason you add a second hire — a real skill or capacity gap, never because a team sounds impressive.