CC Parallel Primitives — Decision Tool

Claude Code gives you three different parallel primitives. They're not interchangeable. This page helps you pick the right one for your workload. Companion to docs/parallel-primitives.md.

Task tool

STAR pattern · subagents · in-process

~150K tok simplest

N independent things in parallel
Parent collects results
No agent-to-agent chat
Dies with parent session

Example: /ork:cover spawns one Task per test tier.

Agent Teams

MESH pattern · TeamCreate + SendMessage

~400K tok debate-capable

Agents react to each other in real time
Design debate, hand-off chains
~3× Task tool cost for back-and-forth
Dies with parent session

Example: /ork:brainstorm with system-designer vs backend-thinker.

claude agents CLI

PROCESS-ISOLATED · separate OS processes

N × budget survives /clear

Overnight / multi-hour batches
Each session has own context budget
Survives terminal close, /clear, reboot
No comms except shared filesystem

Example: M170 wave — 4 detached sessions, one per worktree.

Pick the right primitive

Do the N tasks need to talk to each other while running?

Do they need to survive `/clear`, terminal close, or run overnight?

Q3 (sanity check)

Is the work actually parallel, or is it a single loop over N items?

Case study — today's 5-PR merge queue

Primitive	Fit?	Why
Task tool	❌	Subagents return and die — can't watch persistently
Agent Teams	❌	No coordination needed — rebase is mechanical
`claude agents` CLI	❌	Heavy: N processes for a polling loop is wasteful
A single foreground loop + gh CLI	✅	ONE process, ONE budget, watches N PRs

Lesson: not every "N things" needs parallelism. Sometimes the right answer is a smarter single loop.

🧭 CC Parallel Primitives — Decision Tool

Task tool

Agent Teams

claude agents CLI

Pick the right primitive

Case study — today's 5-PR merge queue