Claude Agent Teams
Anthropic’s subagents — inside Claude Code’s runtime.
Officially “Claude Agent SDK.” Community calls it “Claude Agent Teams.”
Strong tool loop. Flat tree. Heavy token tax.
The three numbers that matter
tokens per subagent call
Context isolation isn’t free.
concurrent maximum
Hard parallelism cap.
level of nesting
Subagents cannot spawn subagents.
Strong vs honest
Strong
✓ Zero-glue tool loop — Read, Write, Bash, MCP built in
✓ Context isolation per subagent
✓ Auth flexibility — API key, Bedrock, Vertex, Azure
Honest
· No agent-to-agent messaging — everything via parent
· TS SDK is commercial-licensed (not OSI-open)
· Single-vendor runtime — Claude Code
When to pick it
When you want Claude Code’s exact tool loop as a library inside your own process and you accept the vendor coupling. Not a replacement for a real orchestrator when you need hierarchy, messaging, or multi-runtime.
docs.claude.com/en/agent-sdk · github.com/anthropics/claude-agent-sdk-python