Skip to content
Call wiki

Can Claude chat or Cowork dispatch parallel subagents?

Last updated 2026-08-17 · Confidence: documented — Cowork’s help center, Claude Code’s subagent docs, and Anthropic’s multi-agent research writeup; the invocation-phrase and how-to-verify claims are inferred.

Cowork can: “sub-agent coordination” is on its documented capability list. Plain chat runs parallel agents only inside the Research toggle.

Cowork: subagents are built in

Section titled “Cowork: subagents are built in”

Cowork’s help center lists sub-agent coordination as a core capability — “Claude breaks complex work into smaller tasks and coordinates parallel workstreams” — and notes that during a task “Claude may coordinate multiple sub-agents working simultaneously.” Nothing needs enabling; Claude decides when to fan out.

There is no magic phrase. An explicit instruction — “collect these with subagents, one per data source” — is just a prompt that makes delegation likely instead of leaving it to Claude’s judgment. In Claude Code, the same pattern is documented: naming a subagent in the prompt makes Claude delegate to it.

What a subagent saves — and costs

Section titled “What a subagent saves — and costs”

Each subagent works in its own context window and returns only a summary; the searching and file-reading never enter the main conversation. That is what keeps a fan-out over many sources from degrading the main session.

The saving is context, not tokens: Anthropic measured multi-agent systems at “about 15× more tokens than chats” — irrelevant on a subscription, real on API billing.

How to tell it really ran

Section titled “How to tell it really ran”

Cowork’s docs promise progress indicators per step, not a subagent badge; parallel dispatch should show as several task lines running at once. The robust check: have each subagent write its output to a file in the session folder, then count the files.

Plain chat: only via Research

Section titled “Plain chat: only via Research”

Chat has no subagent tool, so “dispatch subagents” there yields one model narrating as if it had helpers. The documented multi-agent surface is Research: a lead agent “spawns subagents to explore different aspects simultaneously” and compiles their findings into one cited answer.

Sources

Section titled “Sources”