Anthropic just added Agent Teams to Claude Code (experimental), and it’s one of those features that looks small… until you realise it changes how you work.
Instead of one Claude Code session spawning subagents that only report back, an agent team is multiple Claude Code instances that can:
- run in parallel
- keep their own context
- message each other directly
- coordinate via a shared task list
That’s the missing UX layer for real multi-agent work.
Agent teams vs subagents (what actually changes)
Anthropic’s docs draw the line clearly:
Subagents
- run inside one session
- can only report back to the main agent
- lower token cost
- best for focused “go do X and return result” work
Agent teams
- each teammate is independent (own context)
- teammates can talk to each other
- higher token cost
- best for complex work spanning layers (frontend/backend/tests) or competing hypotheses
Where I’d use agent teams at BuildrLab
1) Cross-layer feature work
- Teammate A: frontend
- Teammate B: backend
- Teammate C: tests/QA
- Lead: integration + final review
2) Debugging with competing hypotheses
Have 2 teammates chase different theories and converge faster.
3) Research + implementation
One teammate reads docs/releases, another codes, another writes tests.
This pairs really well with our existing discipline:
- git worktrees per agent to avoid clobbering
- PRs per unit of work
The tradeoff: token cost vs speed
Agent teams cost more because you’re literally running multiple contexts.
So you only use them when parallelism is real.
If the work is sequential or everyone needs to edit the same file, you’ll waste time and money.
Sources
- Anthropic docs: Agent Teams — https://code.claude.com/docs/en/agent-teams
Top comments (0)