Agentic AI

Agentic AI is AI that carries out a task on its own instead of only answering questions. You give it a goal, such as "submit the receipts from this business trip", and it works out the steps, uses software like your email or expense system, checks each result and asks you when it gets stuck. A program that works this way is called an AI agent.

How does it differ from a chatbot?

A chatbot answers one message at a time (see chatbot vs AI agent). Anthropic describes an agent as a model that decides for itself how to achieve what the user wants, working in a loop: it plans, acts, observes the result, adjusts and repeats until the task is done or it needs input (Trustworthy agents in practice, April 2026). In its example, Claude submits travel receipts one by one. When a hotel charge exceeds the nightly cap, it pauses to ask whether it should fetch the company's expense policy before trying again.

Is every AI system with tools agentic?

It is a matter of degree. In Building effective agents (December 2024), Anthropic calls the whole range agentic systems and draws one line through it. In a workflow, code fixes which steps run in which order, as in prompt chaining. An agent chooses its own steps and tools, which suits open problems where you cannot predict how many steps it will take. Anthropic advises starting with the simplest solution, which may be no agentic system at all: such systems trade speed and cost for better results on complex tasks, and an agent's autonomy means higher costs and a risk of compounding errors.

How does it work?

A large language model does the reasoning. It calls tools, such as a search function, an API or a database, through function calling or the Model Context Protocol (MCP). After each step it reads the result, which Anthropic calls ground truth from the environment, and decides what to do next. A stopping condition, such as a maximum number of steps, keeps the loop under control. Larger tasks can be split: in the orchestrator-workers pattern, one model breaks the task down and hands the parts to other models (multi-agent systems).

What changed in 2026?

In April 2026 three vendors released infrastructure for agents that run in sandboxes or in the cloud, without depending on the user's laptop:

• Anthropic introduced routines in Claude Code (research preview, 14 April): a task you configure once, with a prompt, repositories and connectors, that runs on a schedule, from an API call or on a GitHub event.
• OpenAI updated its Agents SDK (15 April) with native sandbox execution, configurable memory and instructions from AGENTS.md files.
• Cloudflare expanded its Agent Cloud (13 April), with generally available sandboxes and access to OpenAI's GPT-5.4 and Codex, among other things.

The approach also reaches beyond software. In March 2026 researchers at BASF and RWTH Aachen described a system in which one agent solves a chemical process-design problem with engineering knowledge and a second writes the solution as code for BASF's in-house simulation tool (Schäfer et al.).

Where does it go wrong?

• The agent misreads what you want. According to Anthropic, agents act with less human oversight, which leaves more room for actions with unintended consequences.
• Prompt injection: instructions hidden in an email or web page that the agent reads can trick it into costly actions it would otherwise not take.
• The business case. In June 2025 Gartner predicted that over 40% of agentic AI projects will be canceled by the end of 2027, because of rising costs, unclear business value or inadequate risk controls. It also warned of "agent washing", existing chatbots and automation tools rebranded as agents, and estimated that only about 130 of the thousands of agentic AI vendors are real.