Skip to content
JUNG.
bloggalleryplacesguestbookabout
Are Subagents Smarter?

31 Aug 2026 · 3 min read

Are Subagents Smarter?

What I learned from reading about context isolation and delegation

AI

I learned for the first time, while reading an article, that subagents can also help manage context. I have barely used them before, but now I want to try using them more.

I did have one question. Does a subagent know the context that the main agent already knows? If it does not, could the quality of the result suffer?

A separate context from the start

According to the Claude Code documentation (https://code.claude.com/docs/en/sub-agents), a subagent does not continue the main conversation as-is; it starts in a separate context. The task description passed by the main agent becomes an important briefing. There is also information loaded separately at startup, such as configured instructions and skills, so it is not quite accurate to think that a subagent acts from one line of briefing alone.

In the end, what matters is not only what you delegate, but how specifically you communicate the necessary background, the desired result, and the criteria for checking it. Even facts the main agent already knows need to be shared again if the subagent needs them.

What isolation protects, and what it costs

A separate context is not only a disadvantage. During a long task, file contents and tool results keep piling up in the main conversation. Anthropic’s context engineering article (https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents) explains that as information grows, it can become harder to find the signal you need. A subagent keeps the details from its own work inside its context and returns only the key result to the main agent. The article says these summaries are often around 1,000 to 2,000 tokens.

So a subagent does not make the model itself smarter. It is closer to a way of separating the context needed for a task so that the main conversation stays less complicated.

Anthropic’s multi-agent research example (https://www.anthropic.com/engineering/multi-agent-research-system) was interesting too. In its internal research evaluation, the combination of an Opus 4 lead and Sonnet 4 subagents performed 90.2% better than a single Opus 4. At the same time, the multi-agent system used about 15 times as many tokens as a normal chat interaction. That 15x figure is not a direct comparison with a single agent, and the research result cannot be applied to every development task as-is. Nothing is free.

What should I delegate?

The part I need to be careful about seems to be a weak briefing rather than isolation itself. If I delegate work without the necessary background, the subagent spends time rediscovering context the main agent already understood. For work that requires receiving the result and going back and forth several times, it may be better to stay in the main context.

On the other hand, a clearly scoped investigation or independent review is worth delegating. Rather than using subagents automatically, I want to check for myself which kinds of separation actually help.

References