Last updated: 27 September 2026
Short answer: agentic AI is AI that works towards a goal by deciding its own next step, using tools such as search, files, email or code, checking the result, and repeating until it is done. A chatbot answers one message at a time; an agent runs a loop. In 2026 agents are genuinely useful for bounded tasks like coding, research and routine admin, but they still make mistakes that compound over long tasks, and a lot of products labelled "agentic" are ordinary chatbots with new branding.
What is the difference between a chatbot, a workflow and an agent?
The clearest definition I have seen comes from Anthropic's engineering guide, Building effective agents. It separates two kinds of "agentic system":
- Workflows, where AI models and tools follow predefined steps written in code.
- Agents, where the AI model directs its own process and decides which tools to use and when.
Put next to a plain chatbot, the difference looks like this:
| Chatbot | Workflow | Agent | |
|---|---|---|---|
| Who decides the steps? | You, one message at a time | The developer, in advance | The AI, as it goes |
| Uses tools? | Sometimes | Yes, in a fixed order | Yes, chosen at each step |
| Predictability | High | High | Lower |
| Good for | Questions, drafting | Repeatable processes | Open-ended tasks |
| Example | "Summarise this report" | Every invoice is read, checked against the order, then filed | "Find why this test fails and fix it" |
Anthropic's guide also makes a point many vendors skip: the simplest solution that works is usually best, and a workflow is often the right choice before an agent.
How does an AI agent actually work?
Strip away the marketing and almost every agent is the same loop:
- Goal. You give it a task, such as "update the pricing table on our website from this spreadsheet."
- Plan. The model decides a next step: read the spreadsheet.
- Act. It calls a tool, for example a file reader.
- Observe. It reads the tool's result.
- Repeat or stop. It decides the next step based on what it saw, and continues until it believes the goal is met, hits a limit, or asks you for help.
The model itself is the same kind of language model you chat with. What makes it an agent is the loop, the tools it is allowed to use, and the freedom to choose between them.
What can agents reliably do today?
The strongest results are in tasks with clear success criteria and fast feedback:
- Coding. Reading a codebase, making a change, running the tests and fixing failures. Tests tell the agent whether it succeeded, which is why coding agents improved first.
- Research. Searching several sources, reading them and writing a summary with links you can check.
- Routine admin. Filling forms from documents, reconciling lists, drafting replies for a human to approve. Engineering and quality paperwork is a good example, with clear limits, which I covered in how AI is changing engineering documentation and quality work.
Capability is moving quickly. METR, an independent research group, measured how long a task (timed by how long it takes a skilled human) AI agents can complete with 50 percent success. In its March 2025 study, that length had doubled roughly every seven months since 2019, and METR's later updates suggested the pace had picked up. The practical meaning: tasks that take a person minutes are increasingly handled well; tasks that take a person days mostly are not.
What can't agents do well yet?
Long, messy tasks are still hard, for a reason that is easy to show with arithmetic. Suppose an agent gets each individual step right 95 percent of the time, which sounds good. A 20-step task only succeeds if every step does:
0.95 multiplied by itself 20 times is about 0.36.
So roughly a one-in-three chance of a clean finish, even with a strong per-step success rate. Real agents recover from some errors, so it is not quite that bleak, but the principle holds: errors compound. That is why agents do best on short, checkable tasks and worst on long chains with no feedback.
Other current weak spots:
- Ambiguous goals. If "done" is not clearly defined, the agent may stop too early or wander.
- Cost. Every step is another model call. A long agent run can cost far more than a single chat answer.
- Permissions. An agent that can send email or delete files can do damage when it misunderstands.
Why is everyone talking about MCP?
Because agents need a standard way to connect to tools. The Model Context Protocol (MCP) is an open standard that lets an AI application talk to external tools and data sources, such as a file system, a database or a calendar, through one common interface. Anthropic introduced it in November 2024. In December 2025 it was donated to the newly formed Agentic AI Foundation under the Linux Foundation, co-founded by Anthropic, OpenAI and Block.
For a business, the practical upshot is simpler integration. A tool that offers an MCP server can be plugged into many different AI applications instead of needing a separate connector for each one.
Is "agentic AI" mostly hype?
Partly. The technology is real, but the label is overused. In June 2025, Gartner predicted that over 40 percent of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. Gartner also warned about "agent washing", meaning existing chatbots and automation tools rebranded as agents, and estimated that only about 130 of the thousands of vendors claiming agentic AI offer the real thing.
My reading of that is not "avoid agents". It is "be specific". Ask any vendor exactly which tools the agent can use, how it decides, what it costs per task, and what happens when it gets something wrong.
How can a small business try agents safely?
- Start with a workflow. If the steps are always the same, fixed automation is cheaper and more predictable than an agent.
- Pick a small, checkable task. Something where you can see quickly whether it worked.
- Keep a human in the loop. Let the agent draft, and have a person approve anything that goes to a customer, moves money or deletes data.
- Limit permissions. Give the agent access only to what the task needs, preferably read-only at first.
- Set spending caps and keep logs. You want to see every step it took and what it cost.
Clear instructions matter even more for agents than for chat, because the agent acts on them without checking in. The habits in my practical guide to prompt writing apply directly. If data privacy is a concern, local vs cloud AI for a small business covers where your data goes when an AI processes it.
FAQ
Is ChatGPT an AI agent?
The basic chat is not; it answers one message at a time. Several AI products, including ChatGPT and Claude, now include agent features that can browse, run code or use connected tools over multiple steps, which do fit the definition.
Will AI agents replace jobs?
They are automating parts of jobs, especially short, repetitive, checkable tasks. Long, ambiguous work that needs judgement over days is still largely beyond them, and most successful deployments keep a person reviewing the output.
Do I need to know how to code to use an AI agent?
Not necessarily. Many business tools now include agent features with no coding. Building your own agent, or connecting it to your systems, usually does need some technical work.
What is the difference between agentic AI and automation like RPA?
Traditional automation, including robotic process automation (RPA), follows fixed rules that someone wrote in advance. An agent decides its steps as it goes. That makes agents more flexible with messy inputs, and less predictable.
