The Rise of Agentic AI

Most people first met modern AI as a chatbot: you type a question, it replies, and the exchange ends. A newer design is attracting attention, usually called agentic AI. Instead of answering one prompt at a time, an agent is given a goal and decides, step by step, which actions to take to reach it. That sounds like a small change, but it shifts the question from “is the answer right?” to “is the whole chain of actions safe and correct?” This article explains what makes a system agentic, where it is being tried, and where the limits and risks lie, using evidence available up to 1 February 2026.

What makes AI “agentic”

Anthropic, an AI developer, draws a useful line in its December 2024 guide to building agents. Workflows are systems where language models and tools follow predefined code paths. Agents are systems where the model dynamically directs its own process and its use of tools.^[1] In other words, the difference is who decides what happens next: the programmer in advance, or the model as it goes.

The research idea behind this is older than the term. A 2022 paper called ReAct proposed having a language model produce reasoning steps and actions in an interleaved way, so that reasoning guides which action to take and the results of actions feed back into the reasoning. By letting the model query a simple Wikipedia interface, the authors reported reduced hallucination and error propagation on question-answering tasks, and success rates 34 and 10 percentage points higher on two interactive decision-making benchmarks.^[2] The International Federation of Robotics (IFR) uses similar language in its 2026 trends report, describing analytical AI that predicts failures, generative AI that helps robots learn tasks, and agentic AI that combines the two to support independent work in complex real-world settings.^[3]

Where agents are being tried

Software development. Coding is a natural fit because results can be checked by running tests. A benchmark called SWE-bench collects 2,294 real problems from GitHub projects and asks a model to edit a codebase to fix each one. When it was introduced in 2023, the best model tested resolved only about 2% of them.^[4] That figure dates from the benchmark’s launch, and it is the kind of measurement to look for: a defined task, a pass or fail result and a published method.

Healthcare. Claims that AI analyses medical scans with “superhuman precision” deserve care. A 2020 study in Nature compared an AI system with radiologists on breast cancer screening. It found absolute reductions in false positives of 5.7% in a US dataset and 1.2% in a UK dataset, and in false negatives of 9.4% and 2.7%.^[5] That is promising, but it was a retrospective evaluation of an analysis tool, not an agent acting on its own, and the US and UK datasets were built differently. It supports AI as a second reader, not as a replacement for clinical judgement.

Robots and machines. The IFR expects robots that use AI to work independently to become more common, though it is an industry body describing its own expectations.^[3] Our Robotics guide explains how robots sense, decide and act, and Emerging Technologies explains why new capabilities often take years to settle into everyday use.

Risks and limits

Wrong answers. Language models can produce fluent text that is not true, a problem researchers call hallucination.^[6] In a single reply this is a nuisance. In an agent, a mistaken step can feed into the next one, and Anthropic notes that agents bring higher costs and the potential for compounding errors, trading latency and cost for better performance.^[1]

Manipulation. The OWASP Top 10 for LLM applications lists prompt injection as its first risk. In an indirect attack, instructions hidden in a website or file that the model reads change what it does. OWASP lists unauthorised access to the model’s functions and execution of commands in connected systems among the possible impacts, and says severity depends on how much agency the model has.^[7] The more an agent can do, the more it matters what it reads.

Oversight and rules. Frameworks are emerging. NIST’s AI Risk Management Framework, released on 26 January 2023, is voluntary guidance for managing AI risks to people and organisations.^[8] In the European Union, the AI Act entered into force on 1 August 2024, with its obligations applying in stages.^[9] Neither is specific to agents, but both stress identifying and managing risk.

What to take from this

Agentic AI is best understood as a design pattern, not a leap in intelligence. Giving a model tools and a goal can make it useful for multi-step work, and early benchmarks show both progress and a long way to go. It also widens the range of things that can go wrong, which is why testing, limits on what an agent may do, and human review of important actions matter.

When you read a claim about an AI agent, ask what task it was tested on, how success was measured, what the agent was allowed to do, and who checked the result. Our Technology Comparisons and Reflection page has questions for exactly this kind of judgement.