Contents
“AI agent” is one of the most used and least defined terms in technology. Companies apply it to everything from a customer service chatbot to software that writes and tests code for hours without supervision. The clearest way to understand it is to look at what an agent does differently from the chatbot most people already know.
The short definition
A chatbot takes your message and returns a reply. An agent is given a goal and works towards it in steps. At each step it chooses an action, such as searching the web, running code or editing a file. It looks at what came back, then decides what to do next. It repeats this until the job is finished or it gets stuck.
Anthropic, the company that makes the Claude models, draws the line this way in its guide for developers. Workflows are “systems where LLMs and tools are orchestrated through predefined code paths”. Agents are “systems where LLMs dynamically direct their own processes and tool usage”. In a workflow, a programmer decides the order of steps in advance. In an agent, the model decides.
What makes the loop work
Three parts matter.
- Tools. A model on its own can only produce text. Tools let it read files, browse pages, send requests to other software and run programs.
- Feedback. After each action the agent sees a real result: an error message, a search result, a test that passed or failed. Anthropic’s guide calls this getting “ground truth” from the environment. It is what allows an agent to notice and fix its own mistakes.
- Stopping rules. Well-built agents pause for a person at checkpoints, ask before risky actions, and have limits on how long they can run.
How capable are agents today?
The most cited measure comes from METR, a research non-profit. It times how long tasks take skilled humans, then tests which of those tasks AI agents can finish. In March 2025 it reported that the length of task agents could complete half the time had been doubling about every seven months for six years.
The trend has since quickened. In an update in January 2026, using a larger set of 228 tasks, METR estimated a doubling time of about 89 days for models released since 2024. The best model it measured, Claude Opus 4.5, completed tasks that take a human about 320 minutes, more than five hours, half of the time.
Two cautions belong beside those numbers. The first is that METR’s own range for that estimate runs from 170 to 729 minutes, which it describes as “very wide”. The second is that the tasks are mostly software and research work, and the headline is a 50% success rate. In its 2025 study, models succeeded almost every time on tasks that took humans under four minutes and less than 10% of the time on tasks over about four hours.
Succeeding half the time is impressive in a lab and not good enough for a payroll run.”
| Item | Doubling time in days, by period of model releases |
|---|---|
| All models measured | 197 days |
| Models since 2023 | 131 days |
| Models since 2024 | 89 days |
A shorter bar means faster progress.Source: METR, Time Horizon 1.1. Chart: The Sidelight.
Why many agent projects disappoint
Half-right is a poor standard for business work, and that gap shows in adoption. Gartner, the research firm, predicted in June 2025 that more than 40% of agentic AI projects will be cancelled by the end of 2027 because of escalating costs and unclear business value. It also warned of “agent washing”, where vendors rebrand existing assistants and chatbots as agents, and estimated that only about 130 of the thousands of vendors claiming to sell agents offer the real thing.
Interest is wide all the same. In McKinsey’s 2025 global survey, 62% of respondents said their organisations were at least experimenting with AI agents. Gartner itself expects 15% of day-to-day work decisions to be made by agents by 2028, up from none in 2024.
Anthropic’s advice to developers is notably restrained. It recommends “finding the simplest solution possible, and only increasing complexity when needed”, and notes that agents usually cost more and respond more slowly than a single well-designed prompt.
Where agents work well now
Agents do best where three conditions hold: the result can be checked automatically, mistakes are cheap to undo, and the work happens in software.
- Programming. Code either passes its tests or it doesn’t, so the agent gets clear feedback.
- Research. Searching, reading many sources and summarising them, with a person checking the output.
- Routine back-office tasks. Sorting requests, filling forms and moving data between systems, with approval steps for anything that matters.
They do worst where errors are costly and hard to spot, such as moving money, giving medical or legal advice, or acting on someone’s behalf without review.
What to ask before trusting one
- What can it do without asking me, and can I limit that?
- Can I see a record of every action it took?
- What happens when it is wrong, and can the action be undone?
- Has it been tested on tasks like mine, and how often did it succeed?
A vendor who can’t answer these is probably selling a chatbot with a new label.
References and sources
- 01Building effective agents · Anthropic · 19 December 2024
- 02Measuring AI ability to complete long tasks · METR · 19 March 2025
- 03Time Horizon 1.1 · METR · 29 January 2026
- 04Over 40% of agentic AI projects will be scrapped by 2027, Gartner says · Reuters, via The Star · 25 June 2025
- 05The state of AI in 2025: Agents, innovation, and transformation · McKinsey & Company · November 2025
- 06Lead image: Code on computer monitor (Unsplash).jpg · Markus Spiske, via Wikimedia Commons (CC0) · Accessed 13 November 2026
House byline
Articles under this byline are researched and drafted with AI assistance, with every figure checked against its source before publication. The editor, Pankaj Guleria, reviews them and corrects them where needed.
Edited by Pankaj Guleria, Founder and Editor



