Agent Teams (Claude Code)
Claude Code can spawn parallel subagents via the Agent tool: each agent forks a fresh context, works in isolation, and reports back to the orchestrator. Typical patterns: orchestrator + N specialized workers, sequential pipeline, simultaneous multi-perspective analysis.
Strengths
- Native parallelization, N sub-problems → N simultaneous agents, same total duration
- Fresh context per agent, each specialist sees the code without main-session noise
- Multi-perspective analysis in one batch: security + perf + types + accessibility
- Background subagents: four uses that hold up. Priming context (one subagent reads the existing code while others dig through documentation and public implementations), writing tests for a feature while the main session builds it, extracting useful memories from the transcript before hitting the context limit, and watching a local server's logs. That last one is not a subagent: it is a Bash command run in the background, which the agent rereads while you test.
Limitations
- Token cost proportional to agent count, ~3-5× more expensive than solo agent
- Edit conflict risk if agents aren't scoped to distinct files
Best for
- Multi-angle codebase audit (security, perf, quality, tests) in one single run
- Complex feature requiring an orchestrator + several specialized workers