A Claude agent is any application that completes a task by planning its own steps and calling tools. The name covers Claude Code itself, the Claude Agent SDK that exposes the same loop as a Python or TypeScript library, subagents that run inside a session with their own context window, background agents that run many sessions in parallel on your machine, Anthropic-hosted Managed Agents, and the agents in the Claude apps such as Cowork and Dispatch. They are the same loop packaged six ways.
- They all run one loop: read, act, observe, decide, repeat until done.
- Claude Code is the agent most people mean when they say Claude agent.
- The Agent SDK is that same loop as a library, Python and TypeScript only.
- Subagents are inside a session; background agents are separate sessions.
- Anthropic's branding rules allow "Claude Agent" and forbid "Claude Code Agent".
Six things, one name
What people mean by "Claude agent", as of August 2026.
| If you saw | It is | It runs |
|---|---|---|
| Claude Code | The coding agent itself | On your machine, or in a cloud VM |
| Claude Agent SDK | The same loop as a library | In your own process, Python or TypeScript |
| Subagent | A specialised worker inside one session | In its own context window, same session |
| Background agent | Many parallel Claude Code sessions | On your machine, one worktree each |
| Managed Agents | A hosted REST API, a separate product | On Anthropic infrastructure |
| Cowork and Dispatch | Agentic work that is not code | In the Claude apps |
What makes any of them an agent
Anthropic's own definition is short and worth taking literally: an agent is an application that completes a task by planning its own steps and calling tools that read files, run commands, or edit code.
CHATBOT AGENT
you ask you describe an outcome
model answers model reads files
you act on it model runs a command
model reads the output
model decides what next
... until done
you review a diffThe consequential line is the fourth on the right. A model that can observe the result of its own action can run your tests, read the failure, and try again. That feedback loop is the entire reason the category exists, and it is why an agent needs permissions in a way a chatbot never did.
Claude Code: the agent people mean
When a job posting says "experience with Claude agents", this is nearly always what it means. Claude Code is the reference implementation of the loop, and everything else on this page is either it, inside it, or a repackaging of it.
- Tools for reading, editing, searching, running commands, and fetching the web.
- Permission modes that decide what runs without asking, from Manual through Plan to Auto.
- Context management: automatic compaction,
/context, and checkpoints. - Extensibility:
CLAUDE.md, skills, hooks, MCP servers, plugins. - Five front ends: terminal, desktop app, web, mobile, and IDE extensions.
The Agent SDK: the loop as a library
If you want that behaviour inside your own product rather than in a terminal, the Claude Agent SDK gives you the same tools, the same agent loop, and the same context management, callable from your code. It was previously named the Claude Code SDK, which is why half the tutorials on the internet use the old name.
Four Anthropic products that all sound like the same thing.
| If you are | Use |
|---|---|
| Building an agent without writing the tool loop | Agent SDK |
| Doing interactive work or one-off tasks in a terminal | Claude Code CLI |
| Calling the API and writing the tool loop yourself | Client SDK |
| Running long agents without managing a sandbox | Managed Agents |
- Available as a library for Python and TypeScript only. From another language, run the CLI as a subprocess with
-pand--output-format json. - It loads skills, commands, and memory from your project
.claude/and from~/.claude/, exactly as Claude Code does. That surprises people who expected a blank slate. - Authenticate with an API key. Anthropic does not allow third-party developers to offer claude.ai login or subscription rate limits for products built on the SDK without prior approval.
Subagents: agents inside one session
A subagent is a specialised worker that Claude delegates to from inside a session. It runs in its own context window with its own system prompt, its own tool access, and independent permissions, does the work, and returns only a summary.
---
name: db-reviewer
description: Reviews schema and migration changes for safety
tools: Read, Grep, Glob
model: haiku
---
You review database migrations. Flag anything that locks a large table,
drops a column still referenced in code, or lacks a reversible down step.
Report findings only. Do not edit files.- Context preservation is the main reason to use one. A search that would flood your main conversation with thousands of lines happens somewhere else and comes back as a paragraph.
- Constraint is the second. The
toolsfield is a real capability limit, not a suggestion. - Cost is the third. Routing mechanical work to a cheaper, faster model is a per-subagent setting.
- Under auto mode, the classifier checks a subagent three times: the task description before it starts, each action while it runs, and its full action history when it finishes. A
permissionModein the frontmatter is ignored there.
Background agents: many sessions, one screen
Confusingly, claude agents is not about subagents at all. It opens agent view: a single screen for dispatching and monitoring several independent Claude Code sessions running on your machine.
claude agents # open agent view
claude --bg "investigate the flaky test in auth.spec.ts"
claude attach 7c5dcf5d # take over a session
claude logs 7c5dcf5d # recent output
claude stop 7c5dcf5d
claude respawn 7c5dcf5d # restart, conversation intact
claude agents --json --all # machine-readable list| Property | Detail |
|---|---|
| Isolation | Each background session moves into its own worktree under .claude/worktrees/ before editing |
| Turning that off | {"worktree": {"bgIsolation": "none"}} |
| Rate limits | Each session consumes quota independently |
| Lifetime | Local only. They stop on shutdown and survive sleep. |
| Deleting one | Removes its worktree, so commit first |
The agents that are not about code
| Product | What it is | Plans |
|---|---|---|
| Cowork | The agentic surface for knowledge work, in the Claude apps | A tab in the Claude app |
| Dispatch | A persistent conversation in Cowork that routes tasks, and can spawn a Code session on your desktop | Pro and Max only, not Team or Enterprise |
| Managed Agents | A hosted REST API where Anthropic runs the agent and the sandbox | Platform |
| Claude in Slack, Claude Tag | An @Claude that answers in channels and opens PRs | Claude Tag is Team and Enterprise |
Which one you want
| Your question | The answer is |
|---|---|
| How do I use an AI agent on my codebase? | Claude Code |
| How do I put an agent in my product? | Agent SDK, with an API key |
| How do I stop research flooding my context? | Subagents |
| How do I run five tasks at once? | Background agents, or worktrees |
| How do I run agents without hosting a sandbox? | Managed Agents |
| How do I get an agent to do non-code work? | Cowork |
| How do I send an agent a task from my phone? | Dispatch, or Remote Control |
Questions people ask
An application that completes a task by planning its own steps and calling tools that read files, run commands, or edit code. In practice the name covers Claude Code, the Claude Agent SDK, subagents, background agents, Managed Agents, and the agents in the Claude apps.
An agent. It runs a loop: read files, take an action, observe the result, decide the next step. A chatbot answers and stops; an agent can run your tests, read the failure, and try again.
They are the same agent loop. Claude Code is the interactive product; the Agent SDK is that loop as a Python or TypeScript library you call from your own application. Use the CLI to do work, the SDK to build a product.
Yes. It was renamed, and the package and documentation now use Agent SDK. Older tutorials referring to the Claude Code SDK are describing the same library.
A subagent runs inside one session, in its own context window, and returns a summary to the parent conversation. A background agent is a separate Claude Code session with its own worktree, listed in claude agents, and consuming rate limit of its own.
No. Anthropic does not allow third-party developers to offer claude.ai login or subscription rate limits for products built on the Agent SDK without prior approval. Authenticate with an API key instead.
Anthropic's branding guidance allows "Claude Agent", and "Claude" when it sits in a menu already labelled Agents. It does not permit "Claude Code" or "Claude Code Agent", or visual elements that mimic Claude Code.
Python and TypeScript only. To drive the same loop from another language, run the CLI as a subprocess with -p and --output-format json.
Sources
Every figure above was read from these pages on August 2026. Vendors reprice without notice; if you find a stale number, tell us.