A team proposes decomposing a workflow into several specialised agents that coordinate. What justifies this over a single agent with more tools, and what does it cost?

What justifies it

1. Genuinely different capabilities or permissions. An agent that may read customer data and one that may issue refunds should be different principals with different authorisations, because the security boundary is real.

2. Context that does not fit or should not be mixed. If one part of the task requires a large volume of context irrelevant to another, separating them keeps each focused and avoids the degradation that comes from an overloaded context.

3. Different models per subtask. A cheap fast model for classification and routing, a more capable one for the difficult step. This is frequently the strongest economic justification, since most steps are simple.

4. Independent development and evaluation. Separate agents with separate evaluation sets can be improved independently, which matters when different teams own different parts.

5. Parallelism, where subtasks are genuinely independent and latency matters.

What it costs

• Coordination overhead — more model calls, more latency, more tokens spent passing context between agents.
• Error compounding. Each handoff is an opportunity to lose or distort information, and errors accumulate through the chain in a way that is hard to attribute.
• Debugging difficulty. Reconstructing why a multi-agent system produced an output requires tracing across agents, and the reasoning is distributed.
• Non-determinism multiplied. Each agent introduces variance; several in sequence produce behaviour that is hard to reproduce.
• Evaluation complexity. End-to-end evaluation is what matters and is harder to construct, while per-agent evaluation can improve each part while the whole gets worse.

The default worth defending

A single agent with well-designed tools handles more than teams expect. Many multi-agent designs are decomposition by task description rather than by any of the justifications above — a planner, a researcher, a writer — which adds coordination cost for a separation that buys nothing.

The test: does each agent need a different permission set, a different model, or a genuinely separate context? If not, they are one agent with several tools.

The pattern that usually works

An orchestrator with narrow specialised tools, rather than peer agents negotiating. The orchestrator owns the state, the sequence, the budget and the audit trail — the same reasoning that makes orchestration beat choreography in a saga: someone must be able to answer "where is this and why is it stuck".