Building agents: When to fold subagents into your main agent loop

What the article is about

The article is a guide from Claude Platform, an AI building service. It offers a migration kit that helps teams check which subagents could be merged into the main agent. The kit walks through the patterns, compares options, and tests the best one. It also includes a demo that runs on a mock system first.

What subagents are

A subagent is a helper agent with its own memory space (context window), its own tools, and its own instructions. A coordinator controls the helpers. Subagents are useful for checking work, doing tasks at the same time, and limiting who can do what. Many of Claude Platform's own agents use them, including a multi-agent research system.

Key facts from the article

• In several enterprise tests, one agent with skills often beat both a single big prompt and a subagent design on quality.
• Those single agents often cost less and responded faster for each task.
• One customer found that a multi-agent setup ran about 2 times slower than a single agent loop.
• One company had about a dozen specialist subagents, which made its customer agent hard to keep consistent and slow.
• The article suggests that one clear system prompt, 4 to 5 simple data tools, and firm output rules often cover most use cases.

Why it matters

Many companies build a copy of their org chart into their AI systems, with one helper per team. This can cause messy instructions, overlapping helpers, scores that do not show the whole picture, and slow replies. The article says that teams should keep subagents only when they clearly help, and that simpler designs are often better.

What to watch

• Whether more companies test their agents with the migration kit before adding new subagents.
• Whether teams build shared quality tests that cover the whole agent, not just one small task.
• How the trade-off between subagents and single agents changes as models and tools improve.