Subagents or sessions
Claude Code subagents vs parallel sessions
Subagents work for one session and report back. Parallel sessions work independently. Each needs different coordination.
The problem
Subagents and parallel sessions both look like more agents, but they fail differently. Subagents lose detail in their summaries. Parallel sessions can't see each other at all.
With Concord
01
Subagents for side tasks
Use a subagent for focused research or a contained change. It runs in its own context and returns a summary to the session that called it.
02
Sessions for real lanes
Use separate sessions, often in separate worktrees, when pieces of work should progress independently and may outlive one conversation.
03
Coordinate both the same way
A subagent that edits claims a child task under its parent's task. Parallel sessions claim their own tasks. Both show up in one roster.
Result
Pick subagents or sessions for the shape of the work, and still have one view of who owns what.
Common questions
- Can subagents talk to each other?
- Subagents report to the session that started them. Separate sessions, including ones in other harnesses, can message each other directly through Concord.
- Do subagents share the parent's context?
- No. Each subagent has its own context window, which is why anything the parent or other agents need later should be written to the task, not left in the summary.
Further reading
Give your agents shared work-state.
npm install -g @concord-ai/concord-mcp