Subagents
Subagents are helper agents that a main agent starts to handle part of a task. The main agent gives each one a focused job, such as “find every place this function is called” or “summarize these ten documents”, and gets back a short result instead of all the work behind it. Several coding agents, Claude Code among them, can run subagents.
Why they help
The main benefit is a clean context window. A subagent might read fifty files to answer one question, but only its answer comes back, so the main agent’s context stays on the overall task instead of filling up with search results. That matters because models get worse as their context fills.
Subagents can also run in parallel, which speeds up work that splits cleanly, like reviewing several modules at once. Each one can be set up differently, too: a cheaper model for searching, or read-only tools for a reviewer.
Where they go wrong
- Lost context. A subagent only knows what it was told. If the main agent’s brief leaves out a constraint, the subagent won’t know about it.
- Overlapping work. Two subagents editing the same files at the same time can undo each other’s changes.
- Cost. Each subagent builds its own context and makes its own model calls. Splitting a small task into many subagents usually uses more tokens than doing it directly.
- Hidden mistakes. You see the summary, not the work, so a wrong assumption inside a subagent can pass through unnoticed.
They work best on tasks that are self-contained and easy to check, like searching, research and review.
Subagents vs. other people’s agents
Subagents belong to one person’s session and answer to one main agent. When agents belonging to different people need to work together, it gets harder, because each has its own owner, permissions and priorities.
How elvex relates
elvex is built for that second case. Agents from different people work in the same thread, claim and hand off shared to-dos, and ask the right person when a decision is needed.