Claude AI Agents: The Complete 2026 Guide

Last Updated: November 10, 2026

Summary: Claude AI agents are Claude models set up to plan their own steps and use tools, such as files, a terminal, a browser or your business apps, to finish a task with little hand-holding. You can use them without code through Claude Cowork and Claude in Chrome on paid plans, or build your own with Claude Code subagents, the Claude Agent SDK or Claude Managed Agents, which Anthropic hosts for $0.08 per session-hour plus token costs.

Anthropic now offers agents at every level, from no-code tasks in the Claude apps to hosted APIs. This guide maps each route, its cost, the right model, the key safety controls and how Claude compares with OpenAI. Names, prices and limits were checked against Anthropic's official docs and pricing pages on November 10, 2026.

A Claude AI agent is an application where Claude completes a task by deciding its own next steps and calling tools, rather than answering a single prompt. Anthropic's Agent SDK documentation defines an agent as "an application that completes a task by planning its own steps and calling tools that read files, run commands, or edit code."

Unlike a chat, a Claude agent reads the result of each tool call, decides what to do next and keeps going until the job is done, a limit is hit or a human steps in. For the vendor-neutral basics, see our explainer on what AI agents are.

Anthropic offers six routes to a Claude agent in November 2026: Claude Cowork and Claude in Chrome inside the Claude apps, Claude Code with subagents, the Claude Agent SDK, Claude Managed Agents, and the Messages API for teams that write their own loop.

The Model Context Protocol (MCP) runs underneath almost all of these options. MCP connectors are how a Claude agent reaches a CRM, a database or a project tracker, and Cowork, Claude Code, the Agent SDK and Managed Agents all support MCP servers. Our Claude MCP server guide covers that layer.

Non-developers use Claude agents through Claude Cowork, Claude in Chrome and connectors inside the regular Claude apps. Cowork and Claude in Chrome require a paid plan: Pro, Max, Ultra, Team or Enterprise.

Claude Cowork is Anthropic's agent for knowledge work outside coding. Anthropic's Cowork getting-started guide says it "uses the same agentic architecture that powers Claude Code, with no terminal required." According to that guide, Claude Cowork can:

Anthropic warns that multi-step Cowork tasks "use more of your usage than a quick question," because every step consumes tokens.

Claude in Chrome is a browser extension that lets Claude "read, click, and navigate websites alongside you," per Anthropic's Chrome help article. Claude in Chrome works only in Google Chrome and supports recurring browser tasks on a daily, weekly, monthly or annual schedule.

Connectors give Claude agents access to your apps within each service's existing permissions. Anthropic's connectors help page says custom remote MCP connectors work on every plan, though "Free users are limited to one custom connector."

Claude Code subagents are specialized helpers that a Claude Code session hands focused work to. Each subagent "runs in its own context window with a custom system prompt, specific tool access, and independent permissions," according to Anthropic's subagents documentation, and only its result returns to the main conversation.

Claude Code ships with built-in subagents such as Explore, a fast read-only codebase searcher. You add your own as Markdown files with YAML frontmatter in .claude/agents/ for one project or ~/.claude/agents/ for all projects. Only name and description are required, while optional fields such as tools, model and maxTurns restrict access, pick a model and cap run length.

By default, Claude Code runs up to 20 subagents at once, and a subagent can spawn its own subagents up to three layers deep. Anthropic notes that "each subagent spends tokens of its own while it runs," so parallel agents raise usage quickly.

The Claude Agent SDK is a Python and TypeScript library that gives you "the same tools, agent loop, and context management that power Claude Code," per Anthropic. The Agent SDK quickstart lays out the path:

The Agent SDK also includes hooks, subagents, MCP, permissions, resumable sessions, skills and plugins. One rule matters for product builders: unless previously approved, Anthropic "does not allow third party developers to offer claude.ai login or rate limits for their products," so shipped agents use API key billing. Our guide on how to build AI agents covers platform-neutral design.

Claude Managed Agents is Anthropic's hosted agent service, launched in public beta on April 8, 2026. Anthropic's launch post describes it as a suite of composable APIs "for building and deploying cloud-hosted agents at scale," with infrastructure for state, memory, permissions and scheduled execution.

The Managed Agents documentation splits the service into agents (model, prompt, tools, MCP servers and skills, reused by ID), environments (an Anthropic cloud sandbox or your own self-hosted sandbox), sessions (running instances) and events (messages exchanged between your app and the agent). Managed Agents is enabled by default for all Claude API accounts.

The launch post features Notion, Rakuten, Asana, Vibecode, Sentry and Atlassian as early customers and claims teams reach production "10x faster," though the launch post gives no method for that figure.

Regulated businesses should note two limits: Managed Agents is not eligible for Zero Data Retention, because transcripts persist until deleted, and it is not HIPAA eligible.

No-code Claude agents come with a paid Claude plan starting at $20 a month for Pro, while agents built on the API pay per token, plus $0.08 per active session-hour on Managed Agents. Web searches add $10 per 1,000 searches on the API.

Prices come from Anthropic's pricing page, which says usage limits apply to all paid plans. The pricing page names Cowork only in the Team plan list, but Anthropic's Cowork help article, the more specific source, says Cowork is available on Pro, Max, Ultra, Team and Enterprise. Our Claude usage limits guide explains the 5-hour and weekly windows.

Anthropic's API pricing page lists these current token rates per million tokens for agents you build:

Agent tools add overhead: the Bash tool adds 325 input tokens per request on Claude 4.7 and later models, and browser use adds about 6,600. Models from 4.7 onward also use a tokenizer that produces roughly 30% more tokens for the same text. Full rates are in our Claude API pricing breakdown.

Picture a Claude Sonnet 5.5 agent on Managed Agents that runs 30 minutes each workday, reads 200,000 input tokens, writes 20,000 output tokens and runs 20 web searches, with no caching:

That totals $0.84 per run, or $18.48 across 22 workdays. Runtime is the smallest line item, and Anthropic bills it only while a session is running, so caching and cheaper models save more than shorter sessions.

Anthropic recommends starting with Claude Opus 5.5 for most workloads and moving to Claude Fable 5.2 "for demanding reasoning and long-horizon agentic work." Anthropic's models overview positions Claude Sonnet 5.5 as "the best combination of speed and intelligence" and Claude Haiku 5.5 for high-volume tasks "such as classification, extraction, and routing."

All four current models list a 1 million token context window. A sensible pattern is an Opus 5.5 orchestrator with narrow subagents on Sonnet 5.5 or Haiku 5.5. Haiku 5.5 input costs $0.10 per million tokens on shorter prompts, versus $4 for Opus 5.5.

Claude agents rely on permission modes, sandboxes, admin controls and spending caps, but Anthropic says agents with internet access still carry risk, chiefly prompt injection, where malicious instructions hide in content Claude reads.

Anthropic's Cowork safety guide lists the main controls:

The Agent SDK has allow and deny rules, permission modes and hooks that run code before or after tool calls. Managed Agents adds session budgets, a hard dollar cap that pauses a session once its list-price cost reaches the limit, checked between model requests. Team and Enterprise owners can also set each connector action to Always allow, Needs approval or Blocked.

Claude agents fit recurring, multi-step tasks. These illustrative workflows use only documented features:

OpenAI offers a similar two-layer setup. OpenAI's Agents SDK guide says its SDK "runs in your application," while "the Agents API runs a managed harness in OpenAI's service."

The Claude Agent SDK starts from a finished agent, Claude Code, with file, terminal and search tools already wired in, while OpenAI's SDK docs center on agent definitions, handoffs and guardrails. Our roundup of AI agent platforms compares more vendors.

What are Claude AI agents?
Claude AI agents are Claude models that plan their own steps and call tools, such as files, a terminal, a browser or connected apps, to finish a task. You can use them through Claude Cowork and Claude in Chrome or build them with Claude Code, the Claude Agent SDK or Claude Managed Agents.

Can I build a Claude agent without coding?
Yes. Claude Cowork and Claude in Chrome let you describe a task in plain English, schedule it to repeat and connect your apps, with no code. Both require a paid Claude plan, starting with Pro at $20 a month.

Is the Claude Agent SDK free?
The Claude Agent SDK library is free to install for Python and TypeScript, but the agents it runs pay standard Claude API token rates. Anthropic does not allow third-party products built on the SDK to use claude.ai subscription login unless previously approved.

How much does Claude Managed Agents cost?
Claude Managed Agents charges standard Claude API token rates plus $0.08 per session-hour, billed only while a session is running. Web searches inside sessions cost $10 per 1,000 searches.

What is the difference between Claude Code subagents and the Agent SDK?
Claude Code subagents are specialized helpers inside a Claude Code session, each with its own context window, tools and permissions. The Claude Agent SDK is a library that lets you embed the same Claude Code agent loop in your own application.

Claude AI agents now cover every level of the stack, from a no-code Cowork task to a hosted Managed Agents session inside a product. Cowork uses the same agentic architecture as Claude Code, and the Agent SDK packages that agent loop as a library, so skills learned at one level carry over.

Start by automating one recurring task with the simplest option that can do it, set approvals and budgets before widening access, and watch token use.

Sameer's take: I believe AI agents will replace many entry-level office jobs within three years.

⭐ Get more AI Business Weekly in Google: add us as a preferred source on Google.

Written by Sameer Khan with AI assistance, fact-checked against the sources linked above.