AI Agents Explained: From Answers to Actions (2023–2026)

An AI agent is software you give a goal to rather than a line-by-line script: it plans the steps, uses tools such as your inbox or calendar, carries out the task, and reports back when the job is done.

If you run a small business in Greater Manchester and you keep hearing “AI agents” without a straight explanation, this post is for you. By the end you will be able to define an agent in one sentence, see how the technology moved from chat to action between 2023 and 2026, and decide whether one belongs in your week. This is part of the story of AI since ChatGPT.

We are a Manchester technology business that builds AI systems for clients. The chatbot on this site runs on exactly the mechanisms described below: when it books a consultation, it is acting as an agent, not just answering.

What is an AI agent, exactly?

An AI agent is a system that takes a goal, decides which actions to perform, uses tools to perform them, and checks its own progress until the goal is met.

Here is the plainer version. A normal program follows instructions you wrote in advance. An agent is given an outcome instead, such as “follow up every quote that has gone quiet for five days”, and works out the steps itself. The model is the brain; the tools (email, calendar, your database) are the hands.

So what separates an agent from an ordinary chatbot? The chatbot talks. The agent acts. The dividing line is whether the software can reach outside the conversation and change something in the real world.

Most working agents share four parts: a goal, a set of tools they are allowed to use, a loop that lets them check progress and try again, and boundaries that say what they must not do alone. Get those four right and the agent is useful. Get the boundaries wrong and it becomes a liability, which is why the rest of this post keeps returning to them.

How did AI go from answers to actions?

AI went from answers to actions in three stages between 2023 and 2026: first it got tools, then those tools became platforms, then the whole industry reorganised around agents.

When ChatGPT launched in November 2022 it did exactly one thing. It answered. However clever the reply, the model could not check your calendar, send an email, or look anything up. It was, quite literally, all talk.

The table below tracks how that changed, and what each shift meant for a business.

Period	Capability shift	What it means for a business
2023	Function calling lets a model choose actions and fill in their parameters; plugins connect it to browsing and code	A model can finally do something, not just describe it. Booking, lookups and structured tasks become possible
2023–2024	Assistant frameworks and tool standards mature; computer use lets a model operate ordinary software on screen	Agents can use the tools you already run, not only ones built for an API. Far fewer integration barriers
2025	Reasoning-first models and coding agents arrive; building blocks get standardised	Agents that “think before acting” become reliable enough for real, scoped workloads
2026	Agentic roadmaps and very large context windows become the norm	An agent can hold an entire case file or codebase in mind while it works, handling longer end-to-end jobs

Step one: giving the model hands (2023)

The first crack in the “all talk” wall came in 2023, when models gained the ability to use tools.

In March 2023, OpenAI gave ChatGPT plugins: connections to browsing, code execution and third-party services. For the first time a conversation could reach outside itself. You can read OpenAI’s own product history at openai.com.

The quieter, more important breakthrough was function calling, launched in June 2023. Instead of hoping the model would describe what to do, developers could hand it a menu of actions (send_email, look_up_order, book_appointment), and the model would pick one and fill in the parameters reliably, in a machine-readable format.

It sounds mundane. It changed everything. Function calling is the pattern underneath almost every agent running today, including our own site chatbot when it books a consultation.

Step two: assistants became platforms (2023–2024)

Once models could use tools, the labs raced to make tool-connected assistants easy to build.

OpenAI’s Assistants API arrived at its first DevDay in late 2023. Anthropic developed the Model Context Protocol, a standard for connecting models to tools and data, documented at anthropic.com. The surrounding stack of state management, retrieval and evaluation matured from research code into products.

The capability that made people sit up came in October 2024, when Anthropic shipped computer use: a model that could look at a screen, move a cursor, click buttons and type, operating ordinary software the way a person does. AI was no longer limited to tasks someone had built an API for. In principle it could use the tools your business already runs on.

Step three: the agent era proper (2025–2026)

In 2025 agents went from impressive demos to the industry’s organising idea.

Reasoning-first models like DeepSeek-R1, released in January 2025, showed that “thinking before acting” could be trained efficiently. Coding agents moved into real developer workflows, and building blocks were standardised across the major labs. By 2026, agentic roadmaps and million-token context windows meant an agent could hold a whole codebase or case file in mind while working.

The vocabulary changed with it. The question stopped being “what can it write?” and became “what can it do, and can I trust it to do it unattended?”

What do AI agents mean for a Manchester business?

For a small business, an AI agent means handing a repetitive, rule-bound job to software that follows a goal instead of a script.

Think: follow up every quote that has gone quiet, answer enquiries out of hours and book the good ones into your calendar, or pull last month’s numbers and flag anything odd. These are the jobs agents do well in 2026.

Three lessons from the past three years are worth taking seriously.

The reliable wins are narrow. The agents earning their keep are not general-purpose digital employees. They are tightly scoped systems doing one job well: lead follow-up, booking, triage, reporting. Scope is a feature, not a limitation.

Boundaries matter more than capability. Good agent design is mostly deciding what the system may not do without a human: where approval steps sit, when to hand over, what gets logged. That discipline separates a dependable system from an embarrassing one. The UK National Cyber Security Centre publishes sensible guidance on running AI safely at ncsc.gov.uk.

The building blocks are commodity; the assembly is not. Function calling, tool use and computer use are available to everyone. The value is in mapping them onto your actual workflow, which is judgement, not technology. It is the same judgement behind knowing what AI automation can actually do for a given business.

Here is the honest bit. If there is a repetitive job eating your team’s week, the technology to hand it to an agent now exists and is well-proven. The real question is which job to pick first, and that is exactly what our AI automation service works out with you, starting with a free consultation.