# Agents and Claws (/docs/concepts/agents-and-claws)
The bottleneck for most founders and small teams isn't ideas — it's hands. An **agent** is a pair of those hands: a focused AI worker you recruit once, direct in plain language, and reuse across many jobs. You decide the outcome; the agent does the work.
## Agent vs. Claw [#agent-vs-claw]
Two words, one idea seen from two angles:
* **Agent** — the *role*: an AI employee with a name, a mission, instructions, files, and connected channels. This is what you hire, configure, and review.
* **Claw** — the *worker that executes*: the same agent doing real work inside its [cloud computer](/en/docs/concepts/the-cloud-computer). When you watch a Claw run, you're watching your agent read files, browse the web, run code, and commit to Git.
> You manage agents. Claws are how those agents get things done.
## What an agent is made of [#what-an-agent-is-made-of]
| Part | Role |
| -------------- | ------------------------------------------------------------------------------------------- |
| Identity | Name, description, and the job the agent owns |
| Instructions | System prompt, behavior rules, and operating constraints |
| Cloud computer | An isolated, long-running sandbox the Claw works in |
| Drive | The agent's private file workspace and [knowledge base](/en/docs/concepts/drive-and-memory) |
| Skills | Extra [capabilities and SOPs](/en/docs/concepts/skills-and-automations) the agent can call |
| Channels | External surfaces — WhatsApp, Slack, Telegram — bound to the agent |
| Sessions | Independent conversation contexts created as you and others interact |
## Where an agent sits [#where-an-agent-sits]
```text
Space (your company)
└─ Agent (an AI employee)
├─ Cloud computer (where the Claw works)
├─ Drive (private files + knowledge)
├─ Skills (reusable SOPs)
├─ Channels (external entry points)
└─ Sessions (one per task or conversation)
```
A [Space](/en/docs/concepts/spaces-teams-members) can hold many agents, each focused on a different job — support, finance ops, recruiting, research, coding.
## What each agent gives you [#what-each-agent-gives-you]
Recruiting an agent provisions three things:
1. **A dedicated cloud computer** — isolated CPU, RAM, and persistent SSD for that agent.
2. **A human member seat** — invite one teammate to collaborate with the agent.
3. **Monthly credits added to the Space pool** — credits accumulate at the Space level, not as a per-agent cap. See [credits and billing](/en/docs/concepts/credits-and-billing-model).
## When to create a new agent [#when-to-create-a-new-agent]
Create a separate agent when you need a different knowledge base, a different role or tone, different channel routing, or cleaner isolation between teams.
Don't over-split. If it's the same worker doing related jobs (blog, Xiaohongshu, LinkedIn), use multiple [sessions](/en/docs/concepts/sessions-and-tasks) under one agent. Split into a new agent only when identity, permissions, tools, and files all need isolation.
## Good agent design [#good-agent-design]
* Keep one agent focused on one job, with a clear name and mission.
* Put only the files it should actually use in its Drive.
* Install Skills deliberately instead of turning everything on.
* Connect channels after the core behavior is stable.
## Related [#related]