Agents & Automation

What this section covers

An agent is a program in which a language model does not just answer but decides what to do next: it picks a tool, uses it, looks at the result and repeats until the task is done or it needs your input. The guides in this section explain what that means, how agents connect to tools and data, and what they cost and risk in a company. They rely on the vendors' own documentation, on the Model Context Protocol specification and on the OWASP list of risks for language model applications, and every figure carries the date we checked it.

Where to start

1. What is an agent? gives the definitions from Google Cloud and Anthropic, the difference between an agent, an assistant and a bot, the plan, act and observe loop, real examples and the two OWASP risks that apply to agents directly: prompt injection and excessive agency.
2. What is MCP? explains the Model Context Protocol, the open standard that lets one tool server work with any compatible application. It covers hosts, clients and servers, the three things a server can offer (tools, resources and prompts) and what the specification says about consent and tool safety, including that the protocol itself cannot enforce those principles.
3. agents for business turns the concepts into decisions: which tasks suit an agent, whether you need one at all, three ways to get one, a worked cost example for a support agent handling 5,000 conversations a month, the risks, the EU rules and a 30-day pilot.

• n8n vs Zapier vs Make compares the three automation platforms by how they bill (executions, tasks or credits), prices the same workflow on each, and covers their agents, MCP support and self-hosting.
• How to build an agent goes from deciding whether you need an agent to a tested 60-line example with two tools, a step limit, rules enforced in code and human approval for refunds, with the monthly cost on four models.
• Function calling explains how a model asks your code to run a tool and how OpenAI, Claude and Gemini handle the request, the tool result and strict schemas.
• How to Add an MCP Server: Claude, ChatGPT and VS Code walks through adding local and remote MCP servers to Claude, Claude Code, VS Code and ChatGPT, and the security checks to do first.
• Self-Host n8n: Docker Setup, License and Free Limits covers Docker and Compose setup, what the free Community edition leaves out and what the Sustainable Use License allows.

Do you need an agent at all?

Often not, and the best source for that advice is the vendor that builds agents. Anthropic's guidance is to find the simplest solution possible and add complexity only when needed, which "might mean not building agentic systems at all". In order of complexity: a single well-designed request to a model with the right documents attached, then a workflow of fixed steps with a model doing some of them, and only then an agent that decides its own steps. Agents trade latency and cost for better results on open-ended tasks, so they earn their place when the path to the answer varies too much to script. Our business guide gives the questions to ask before you build.

What an agent costs

The model calls are the easiest cost to estimate. In our worked example, a support agent making six model calls per conversation costs from $17 to $2,075 a month depending on the model, for the same 5,000 conversations. Two conclusions hold whatever your figures are: the choice of model moves the bill by two orders of magnitude, and model usage is often not the biggest cost, because building integrations, testing on real cases, monitoring and review time can each cost more than the tokens. Because an agent repeats the same long prefix at every step, prompt caching matters more for agents than for chat, and our LLM API cost calculator lets you test your own numbers.

What can go wrong

The OWASP Top 10 for LLM applications (2025) names the risks that matter for agents: prompt injection, sensitive information disclosure, excessive agency and unbounded consumption. Anything an agent reads, such as a web page, an email or a document, can try to give it instructions, and an agent that holds more permissions than its task needs can do real damage. The practical guardrails are the same in every guide: narrow permissions, a person approving irreversible or costly actions, a maximum number of steps, a log of every action and regular review of a sample. For why a fluent answer is not a reliable one, see model hallucinations.

If you operate in the EU

Regulation (EU) 2024/1689 already applies to companies that use language models, not only to those that build them. Since February 2, 2025 providers and deployers have had to take measures to support the staff literacy of the staff who use these systems, and from August 2, 2026 the providers of systems designed to interact directly with people, such as chatbots, must make sure those people know they are dealing with a machine, unless it is obvious. Our guides explain the full timeline in the Regulation (EU) 2024/1689 guide and the transparency duties in Article 50 of Regulation (EU) 2024/1689.

What to read next

If you want an agent that writes code, see how to use Claude Code and best coding assistant. If you want to give a model your documents, compare what RAG is with fine-tuning.