Subagent

What it means

The parent agent delegates a scoped goal to a subagent with its own tools, context limit, and completion criteria, then merges results or waits for approval.

Why designers should care

Subagent UX needs spawn status, scope labels, cost visibility, and a single approval surface—users should not chase three hidden chats to understand one task.

Example

A parent “Ship feature” agent spawns a subagent to run tests; the UI shows “Subagent: running vitest (2/4 suites)” with Expand logs and Stop child only.

Common mistakes

• • Subagents that act with broader permissions than the parent.
• • No way to stop a child without killing the entire parent run.
• • Exposing raw subagent chatter instead of a summarized handoff packet.

Related terms

• A2A (Agent-to-Agent): Agents pass messages, goals, or structured handoffs to peers (planner → researcher → writer) often via protocols, queues, or orchestration layers like MCP or custom APIs.
• Orchestration: An orchestrator decides what runs next: call search, pass results to the LLM, invoke a tool, branch on validation, or pause for approval.
• Claw: The claw metaphor describes how autonomous layers extend from chat into systems of record. Users feel it as background work, spawned sub-agents, and actions outside the main conversation thread.
• Background Agent: The agent session decouples from the foreground UI: work continues with its own logs, tools, and time limits until done, failed, or cancelled.
• Multitask: The product queues or runs concurrent agent sessions, each with its own context, tools, and progress, while the user continues working elsewhere.