Key takeaways

• A subagent is a fresh agent with its own context window; its main purpose is context isolation, keeping noisy intermediate work out of the parent's window.
• Built-in subagents (Explore, Bash, Browser) isolate the noisiest operations; Explore uses a cheaper model to run many searches in parallel (tiered execution).
• Run subagents in foreground (blocking, you need the result) or background (parallel workstreams); give each a clear contract: self-contained input, structured output.
• The orchestrator pattern chains specialists (planner, implementer, verifier); the verifier is a skeptical agent that catches work falsely marked done.
• Subagents cost more tokens and add startup overhead; their benefit is isolation and parallelism, not speed. Keep them few and sharply described.

Original sources: Cursor's Subagents docs, Anthropic's How Claude Code works, the "Inside the Codex Agent Loop" deep-dive, and the "Seven Pillars" breakdown.

---