Guide
Anthropic AI agents: products and building blocks
Harrowtech.ai · last checked 13 October 2026
Anthropic does not sell one thing called an agent. It offers several routes to AI agents with Anthropic's Claude models, and they differ in who runs the agent and how much you build yourself. This page sets out the routes in Anthropic's own words, then the advice Anthropic gives to anyone building agents.
Harrowtech.ai is independent of Anthropic. We do not repeat the definition of an agent here; AI agents vs workflows vs agentic AI covers it with Anthropic's own wording.
The four routes
Anthropic's Agent SDK page compares them in a table and says they “differ in who runs the agent, what comes built in, and how you reach it.” This is a plain-English version of that table.
| Claude Code CLI | The terminal interface, for interactive development and one-off tasks. Anthropic's overview calls it “an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools.” It also runs in IDE extensions, a desktop app and the web. |
|---|---|
| Agent SDK | For embedding Claude Code's agent in your own Python or TypeScript application, running in a process you operate. It is a library that runs the Claude Code binary, with its built-in tools, permissions, sessions and hooks. |
| Client SDK | Direct access to the Claude API from your own code. You write the tool loop yourself, or let the client SDK's beta tool runner drive it. |
| Managed Agents | Anthropic hosts the agent harness, which runs the agent loop, and you configure it through the Claude API. Sessions run in an Anthropic-managed cloud sandbox or a self-hosted sandbox on your own infrastructure. You reach it from the SDK for your language, the ant CLI or the REST API. |
Claude Code as an agent
Anthropic's overview lists what Claude Code does beyond chat. It works with git: it can stage changes, write commit messages, create branches and open pull requests. It can spawn several agents that work on parts of a task at the same time, with a lead agent coordinating. In CI, Anthropic says you can automate code review and issue triage with GitHub Actions or GitLab CI/CD. You can also schedule it: the docs describe routines that run in the cloud, so they keep running when your computer is off, and can trigger on API calls or GitHub events. Our Claude Code subagents guide covers the helper agents inside it, and Claude Code pricing covers what plans cost.
The Agent SDK
Anthropic's subtitle for it is “Build production AI agents with Claude Code as a library.” Its documentation says the SDK gives you “the same tools, agent loop, and context management that power Claude Code, programmable in Python and TypeScript.” The capabilities it lists:
- Built-in tools to read, write and edit files, run commands and search the web.
- Hooks, which run your own code at set points in the agent's work.
- Subagents for focused subtasks.
- MCP, the open standard for connecting external tools and data.
- Permissions, to decide which tools run on their own and which need approval.
- Sessions, which keep context and can be resumed or forked.
- Skills, commands, memory and plugins.
Two rules from the same page are easy to miss. Anthropic does not allow third-party developers to offer claude.ai login or rate limits in their products, including agents built on the SDK, unless previously approved; you use API key authentication instead. And the SDK is governed by Anthropic's Commercial Terms of Service. The branding rules say your product should not be called “Claude Code” or “Claude Code Agent”.
Managed Agents
Anthropic describes it as a “pre-built, configurable agent harness that runs in managed infrastructure”, best for “long-running tasks and asynchronous work.” Instead of building your own agent loop, you define four things:
- The agent: model, system prompt, tools, MCP servers and skills.
- The environment: an Anthropic-managed cloud sandbox, or a self-hosted sandbox on your own infrastructure.
- A session: one running instance of the agent doing a task.
- Events: the messages passing between your application and the agent.
Its built-in tools are bash, file operations, web search and fetch, and MCP servers. The docs also mention scheduled deployments for recurring runs on a cron schedule.
Three cautions from the documentation. It is in beta, and every request needs the beta header managed-agents-2026-04-01. Access is “enabled by default for all API accounts”, although “MCP tunnels and dreaming are in a more limited research preview” that you request access to. And because sessions are stored on Anthropic's side, Managed Agents is “not currently eligible for Zero Data Retention or HIPAA Business Associate Agreement (BAA) coverage.” If you handle personal data, read that before you build. Anthropic’s Managed Agents overview shows no price.
Anthropic's advice on building agents
The main source is “Building effective agents”, published on 19 December 2024. It is older than the products above, and it says itself that much of the tooling it describes has since changed and points to the Managed Agents documentation. Three parts are worth knowing.
Three principles. Keep the design simple. Show the agent's planning steps so people can follow what it is doing. And treat the tools you give the agent as an interface that needs real design and testing, which the post calls the agent-computer interface. In its SWE-bench appendix, Anthropic says its team spent more time optimising tools than the overall prompt.
Caution about frameworks. The post says frameworks, including the Claude Agent SDK, can speed up getting started, but add layers that can hide prompts and responses and make debugging harder. It suggests starting with the LLM APIs directly. It adds that frameworks “can also make it tempting to add complexity when a simpler setup would suffice.”
Five workflow patterns. Prompt chaining splits a task into a sequence of calls. Routing sends each input to a specialised follow-up. Parallelisation runs calls at once, either splitting a task or voting across attempts. Orchestrator-workers has a central model delegate to others. Evaluator-optimiser has one model draft and another critique in a loop.
Anthropic names two places where agents earn their cost. Customer support, because it combines conversation with actions such as refunds or ticket updates, and success is easy to measure. And coding, because automated tests can verify the result. Even there, the post says human review stays important.
What this guide does not cover
It does not cover prices; API token prices are in Claude Code pricing. It does not compare Anthropic's tools with other vendors', except in our Claude Code alternatives guide and the GitHub side in AI agents on GitHub. It does not teach you to write code with the SDK.
Where Harrowtech fits
Harrowtech.ai offers AI setup for businesses from £720, including AI agents installed, connected to your systems and locked down, and independent testing of an agent you already have, with a written report. See all Harrowtech guides.
Sources, all read 13 October 2026
- Agent SDK overview, Anthropic
- Claude Managed Agents overview, Anthropic
- Claude Code overview, Anthropic
- Building effective agents, Anthropic, 19 December 2024
Related: AI agents vs workflows vs agentic AI · AI agent architecture: the parts and how they fit · Claude Code subagents explained · All guides