claude agents: run multiple tasks in parallel from the terminal
In brief
Anthropic's new `claude agents` CLI mode lets you fire off independent tasks simultaneously and navigate between them — turning Claude Code into a proper multi-agent terminal environment.
Contents
For most of Claude Code's life, the terminal interface was linear. You ran claude, you had a conversation that worked downward, you waited for one task to finish before starting another. It was powerful — but one lane.
That changed with claude agents.
What it is
Type claude agents in your terminal instead of claude and you enter a different mode entirely. Instead of one chat, you get a panel where each task you submit runs as an independent agent. You can have five things running at once.
The navigation is keyboard-driven: arrow keys to scroll between agents, right to open a chat and see what it's doing, left to go back to the overview. Each agent shows its working status — you can glance across all of them like a dashboard.
Why this matters
The practical implication: you can now run parallel research, parallel builds, or any set of tasks that don't depend on each other — without waiting.
Old workflow: submit task → wait → review → submit next task.
New workflow: submit five tasks → while they run, check in on each → review results as they complete.
For anyone running Claude Code seriously — research pipelines, multi-step dev tasks, content generation at scale — this is a meaningful throughput change. The bottleneck shifts from "waiting for Claude" to "how fast can you review and redirect."
How it works with sub-agents
If you've already built custom sub-agents (specialized agents for specific research domains, writing tasks, coding patterns), they work inside claude agents exactly as before. Each agent session can use whatever sub-agents you've wired up. claude agents is the orchestration layer on top: you're dispatching tasks, not micromanaging how each one runs.
What it's not
This isn't the same as Claude Code's desktop parallel sessions sidebar — that's a GUI feature for running multiple project sessions side-by-side in the desktop app. claude agents is the terminal-native equivalent: same idea, different surface, no GUI required.
If you're running Claude Code headless, in SSH sessions, or just prefer the terminal, claude agents is how you get parallel work without switching to the desktop app.
A note on cost
Each agent in claude agents runs its own independent context window — five parallel agents means five separate context windows running concurrently. Your token usage scales with the number of active agents. For teams monitoring API costs, treat each agent session as a separate call stream: if you're on a usage-based plan, parallel agents will increase spend at a proportional rate. The efficiency gain is throughput, not token reduction.
Getting started
If you have Claude Code installed and updated:
claude agents
From there, type a task and press enter. It queues as an agent. Add more tasks — they all run concurrently. Use arrow keys to navigate. Right arrow on any agent opens it so you can see its progress or add a follow-up.
The most useful starting pattern: fire off 3–5 independent research or build tasks while you work on something else. Check back when they complete.
For the broader Claude Code setup — CLAUDE.md, hooks, project structure — Claude Code project setup covers the foundation. For knowing when to use agent teams vs. sequential tasks, when to use Claude Code agent teams has the decision framework. For the desktop parallel sessions experience, Claude Code parallel agents covers the GUI version.