Agentic AI: systems that plan and act on their own

The word agent is attached to almost every AI product today, which makes it easy to lose sight of what it means. Agentic AI describes systems in which a language model does not just answer but directs its own process: it breaks a goal into steps, chooses and uses tools, looks at the results and decides what to do next.

Workflows and agents

Anthropic draws the line clearly in its guide Building effective agents. Workflows are systems in which language models and tools are orchestrated through predefined code paths. Agents are systems in which the language model dynamically directs its own process and tool usage, keeping control over how it accomplishes the task.

	Workflow	Agent
Path	fixed in advance	decided while working
Predictability	high	lower
Suited to	well-understood, repeatable tasks	open-ended tasks with unpredictable steps
Cost per task	lower	higher, because of more model calls

Anthropic recommends finding the simplest solution possible and only increasing complexity when it is needed. According to the same guide, the basic building block of agentic systems is a language model enhanced with retrieval, tools and memory; tools are increasingly connected through open standards such as the Model Context Protocol.

Where agents help today

From its work with customers, Anthropic names two particularly promising applications: customer support and coding. A support agent can look up an order, check the return policy, draft a reply and hand the case to a person when it is unusual. A coding agent works through a repository much as a developer would, editing files and running tests until they pass; how such an agent hands parts of the job to helpers is shown in our guide to Claude Code subagents.

Five questions before you deploy an agent

• What is the worst action it could take? If it could transfer money, cancel a contract or publish something, a person approves that step before it happens.
• What does it need? Start from no access and add each permission the task proves it needs, read access first and write access only where it cannot be avoided.
• When does it give up? Decide in advance how many steps and how much spend a task may take, and what happens when either limit is reached.
• What does it read that someone else wrote? Web pages, emails and documents can carry hidden instructions, and red teaming tests whether the agent follows them.
• How will you know it went wrong? Log each step, so that a bad outcome can be traced to the step that caused it.

Agents under the EU AI Act

One obligation applies from the first day employees work with an agent: under Article 4, providers and deployers take measures that support the AI literacy of their staff. Our staff training under the EU AI Act is built around it.

Everything beyond that depends on the use case rather than on the technology, because the regulation’s definitions do not treat agents as a category of their own. An agent that filters job applications falls into an area the Act lists as high-risk; one that sorts support tickets usually does not. Deciding where an agent would genuinely help is the starting point of an AI implementation.

Related terms

Term	What it means
Tool use	A model calling external functions, such as a search or a database query, with structured arguments.
Human in the loop	A person who reviews or approves an agent’s actions at defined points.
Embedding	The basis of the retrieval many agents use to find information.

Sources

• Anthropic, “Building effective agents”, 19 December 2024, accessed 14 September 2026
• Regulation (EU) 2024/1689, Annex III, accessed 14 September 2026

Legal status: 4 September 2026. These notes are an editorial summary and not legal advice; they do not replace a lawyer’s review of your particular case.

This page expands an entry from the minoka AI glossary, which covers many more terms in brief.