Tag: subagents

The cold start of a subagent is not what costs you money. Measured across 117 real Claude Code subagent transcripts, startup context was 0.6% to 18% of each agent's total spend. Depth is expensive because every layer runs its own agentic loop and hands the layer above a summary instead of evidence. Here is the compression test that tells you which layer to delete.

A named subagent starts cold, a fork inherits your whole conversation and shares your prompt cache, and a background agent is a second Claude Code session entirely. The trap is the naming: /fork stopped being the in-session fork at v2.1.212 and /subtask took over.

By default a headless Claude Code run emits only tool_use and tool_result blocks from its subagents, so the delegated reasoning is invisible. Pass --forward-subagent-text (Claude Code 2.1.211+) to get subagent text and thinking in the stream-json output, and demultiplex it by parent_tool_use_id.

Version 2.1.219 raises the default subagent spawn depth from 1 to 3, adds a workflowSizeGuideline settings key, and ships a fail-closed sandbox network allowlist.

Version 2.1.218 moves /code-review off your main conversation into a background subagent, and fork-context skills now background by default. Here is what changed and how to opt out.

Version 2.1.213 caps concurrent subagents and stops nested spawns by default, building on the per-session limits from 2.1.212. Here are the new defaults and env vars.

Both let one agent spawn isolated workers from a Markdown file, but Claude Code nests five levels deep with per-agent tool allowlists and worktree isolation, while Cursor keeps it to two levels with a single readonly switch and reads Claude's format for free. Pick by where your agents run, not by brand.

Build a skill to change how Claude works, a subagent to protect your context window, and an MCP server to reach a system Claude cannot otherwise touch. They solve three different problems, not one. Here is the decision, the mechanics, and the config for each.

Define named subagents on agents in the Cursor SDK, let a code-reviewer delegate to a test-writer through the built-in Agent tool, and understand the real nesting depth cap: a subagent spawned by another subagent cannot spawn further ones, despite the changelog's 'and so on'.

Claude subagents give you orchestration-level parallelism out of the box: one agent spawns isolated sub-contexts that run concurrently. OpenAI Assistants never had that, and it shuts down August 26, 2026, so new parallel work belongs on the Responses API plus your own fan-out. Here is how the two models actually differ and which to reach for.