Multi-agent Claude systems: when a second agent earns its keep

A multi-agent Claude system is more than one model loop coordinating: a researcher, a writer, a reviewer, a specialist on SAP. It is justified when a single agent with a small tool set cannot keep permissions or context straight. It is not justified because a diagram looked impressive.

Start single. Add a second agent when you can name the seam.

Seams that are real

• Read-only research versus mutating a system of record
• A specialist that owns one MCP server
• A reviewer agent that only sees the proposed write and the policy

Seams that are fake: "planner, critic, executor, reflector" on a six-step ticket with one API.

What you take on

Every extra agent needs traces, budgets, evals and an owner. Failures become "who dropped the handoff." If you do not have CI eval gates for one agent, you will not have them for four.

Orchestration is still your code

Do not hide the graph inside an opaque vendor workflow if you need to debug a wrong write at 2 a.m. Keep the graph in the repository. That matches how we architect production agents.

When the seam is real, multi-agent is how you keep a write-capable loop from reading the entire internet into the same context window. When it is not, it is how pilots stall with more moving parts. We will tell you which one you have on an audit.