What is Claude Code? The AI agent in the terminal, simply explained

The shortest answer: Claude Code is an AI agent from Anthropic that does not answer, it works. You give it a task, it makes a plan, reads your files, runs commands, checks its own result and delivers. That sounds like a detail, but it is the biggest shift since ChatGPT launched: away from question-and-answer, towards delegating real work.

I (Sebastian) now explain Claude Code in workshops to managing directors and business teams, not just developers. Not because everyone should work in a terminal, but because Claude Code is the blueprint for what agentic work will look like everywhere. Once you understand the principle, you also understand Claude Cowork, the variant for everyone without code.

As of October 2026

Chatbot vs. agent: the difference in one sentence

A chatbot gives you text that you have to implement yourself. An agent implements. Claude Code runs in the terminal (by now also as an extension for the common development environments, in the Claude desktop app and in the browser), has tools like reading files, editing files, running commands and starting tests, and uses them in a loop: plan, execute, check the result, improve. Exactly this loop, the agent loop, is what separates agentic AI from classic AI. We wrote up the basics in the executive crash course on agentic AI.

The thinking behind it is done by Anthropic's Claude models. The default model in Claude Code is Claude Opus 5.5, released by Anthropic on 22 September 2026, on the Pro, Max, Team and Enterprise plans and through the Anthropic API. Since 28 September 2026 Claude Sonnet 5.5 has been available as the lower-cost model. You switch yourself with the /model command, and companies can restrict the choice centrally (more on that below).

What Claude Code actually takes over

Four examples from real projects, small to large:

The bug fix. "The total in the invoice view is wrong." Claude Code finds the spot in the code, identifies the cause, writes the fix, runs the tests and summarises what it changed.

The test coverage. "Write tests for the order module." The agent reads the module, recognises the edge cases and produces a test suite that a human reviews.

The migration. "Move these 40 files from framework A to B." Mind-numbing busywork that used to eat days runs through as a supervised agent task.

The documentation. "Explain the payment flow and write it up as documentation." The agent reads its way through unfamiliar code faster than any new hire.

The common denominator: these are closed tasks with a checkable result, not conversations.

What Claude Code is not

Claude Code is no autopilot. Three things belong to the honest answer:

It needs context. Without a CLAUDE.md that gives the agent your architecture, conventions and no-gos, it guesses generically. How to layer that context is in our three-layer architecture.

It needs boundaries. Through permissions you define what the agent may do alone, what it must ask about and what is off-limits. That is not distrust, that is operational safety. Companies set these rules centrally through managed settings, on the Team and Enterprise plans also directly from the claude.ai admin console. Since late September 2026 they can also pin down models and providers: deniedModels blocks individual models, availableModelsMatch set to "exact" allows only the exact model version that was approved, so new models stay blocked until they are approved, and allowedProviders limits which API providers a machine may use. Project settings can no longer loosen or turn off a sandbox that the admin requires.

It needs review. The agent checks its own result, but responsibility stays with the human. Who is liable when AI gets it wrong is covered in our liability analysis.

What it costs and how it handles your data

Depending on the setup, Claude Code runs through a Claude subscription (Pro, Max), through a Team or Enterprise plan, through the Claude API billed by token (an account in the Claude Console) or through cloud providers such as Amazon Bedrock, Google Cloud and Microsoft Foundry (as of October 2026). On the Team plan, Claude Code is included in every seat, and premium seats add more usage for heavy users. On new Enterprise plans Claude Code is part of the regular seat, while older Enterprise contracts need a seat type that includes it. On Pro and Max, Claude and Claude Code share the same usage limits. Scripts using claude -p also still draw from the limits of the respective subscription: Anthropic paused a change announced for 15 June 2026 that would have moved this usage to a separate monthly credit.

If you pay by token through the Claude API, Opus 5.5 costs 4 US dollars per million input tokens and 20 US dollars per million output tokens, Sonnet 5.5 costs 2 and 10 US dollars respectively (as of October 2026).

Important for companies: on Team, Enterprise, the API and cloud providers, Anthropic states in its documentation that it does not train models on code or prompts sent to Claude Code, unless the customer explicitly opts in. On the consumer plans Pro and Max, a privacy setting in the account decides whether your data is used for training. It has to be checked actively. The big vendor comparison is in the enterprise comparison of ChatGPT, Copilot, Claude.

Why this matters even without a developer team

The principle behind Claude Code, delegate a task, the agent works with tools, a human checks the result, is not limited to code. With Claude Cowork, Anthropic brings exactly this way of working into everyday office life, no terminal required. If you want to know what both feel like in practice, the scenario is in a workday with AI agents.

FAQ

Do I need to code to understand Claude Code? No. To operate it, coding helps. But the principle (delegate, tools, check) is the actual lesson, and it applies to every department.

Does Claude Code replace developers? It shifts their work. Routine moves to the agent; architecture, review and priorities stay human. How much that delivers depends on how carefully context, boundaries and review are set up.

How do you roll it out in a team? Not by distributing licences, but with configuration and training: CLAUDE.md, skills, permissions, a measurement system. The complete playbook is under rolling out Claude Code in the team.

Is it safe for sensitive code? The listed plans (Team, Enterprise, API or a cloud provider) and centrally managed permissions are important safeguards. IT must assess whether Claude Code is suitable for your sensitive code based on data flows, storage and the specific configuration. The personal consumer plans do not belong in a company rollout.

Read more

• Claude Cowork explained: the AI agent for everyone who does not code
• A workday with AI agents: Claude Code and Cowork in the Mittelstand
• What is agentic AI? The executive crash course
• CLAUDE.md, Skills and the three-layer architecture
• Rolling out Claude Code in the team: training, configuration, governance
• Cursor vs GitHub Copilot vs Claude Code compared
• How AI agents work (loop, tools, memory, planning)

Sources

• Anthropic, Claude Code Docs, Overview (surfaces, CLAUDE.md)
• Anthropic, Claude Opus 5.5 (release 22 September 2026, API pricing)
• Anthropic, Claude Code changelog (versions 2.1.280, 2.1.283, 2.1.284, 2.1.285)
• Anthropic, Claude Code Docs, Model configuration (default model per plan, /model, model restrictions)
• Anthropic, Claude Code Docs, Authentication (plans and providers)
• Anthropic, Claude Code Docs, Server-managed settings
• Anthropic, Claude Code Docs, Data usage (training use per plan)
• Anthropic Help Center, Use Claude Code with your Team or Enterprise plan
• Anthropic Help Center, Use Claude Code with your Pro or Max plan
• Anthropic Help Center, Use the Claude Agent SDK with your Claude plan (update of 15 June 2026)

How Sentient Dynamics can help

Our Claude Code course takes you in 22 lessons across 7 modules from tooling setup to multi-agent orchestration. It is part of our licence: €599 per licence, one-time, training plus Cowork and Claude Code. So the no-code entry point with Cowork is included too. Whether it fits your team, we clarify in a 30-minute call.

Book a demo