The Daily Byte
Breaking
Welcome to The Daily Byte — real stories from across Pakistan.
Saturday, 21 November 2026 · Pakistan

The Daily Byte

Stories from Pakistan that deserve to be heard
Minor World

What Is an AI Agent? Apple Mac Plan Explained

By The Daily Byte ·20 Nov 2026 ·10 min read
Aerial view of Apple Park, Apple’s circular headquarters in Cupertino, California.

An AI agent is software that does a job for you, rather than just answering a question. You give it a goal, and it works out the steps, uses the tools it needs, and carries the task through — browsing the web, opening apps, running commands, writing code and calling online services along the way. That is the definition IBM gives: an AI agent is an autonomous AI program that can perform tasks and accomplish goals on behalf of a user or another system without human intervention, by designing its own workflow and using available tools.

The idea moved to the centre of the technology news this week, after reports that Apple is building its own AI agent that can use a Mac in the way a person does. A job listing Apple posted on 17 November 2026, spotted by Bloomberg and reported by TechCrunch on 19 November, describes exactly that: “Imagine an AI agent that can use a computer the way a capable person does. It browses the web, works across apps, runs commands, writes and debugs code, and calls services, and it chooses the right tool for each step.” This explainer sets out what an AI agent is, how it works, how it differs from a chatbot, what Apple is planning, and what the risks are.

What is an AI agent, in plain terms?

A chatbot waits for a prompt and then produces a reply. An AI agent starts with a goal and keeps going until the goal is done, or until it needs your help. If you ask a chatbot to summarise a document, it summarises it. If you ask an agent to prepare a trip, it might check your calendar, search for flights, compare prices, book a hotel, and then show you the result — deciding for itself which tools to use and in what order. The key words in IBM’s definition are autonomy and workflow: the agent designs its own plan and picks its own tools, instead of following a fixed script written in advance.

Agents are usually built on top of large language models, the same technology behind chatbots, but with extra parts bolted on. The model provides the reasoning and language; the agent adds memory, so it can remember what it has already done; planning, so it can break a big goal into smaller steps; and tool use, so it can act in the world rather than only talk about it. IBM describes the difference as one of agency: unlike chatbots and other AI models, which operate within predefined limits and need human input at each step, agents show autonomy, goal-driven behaviour and the ability to adapt when circumstances change.

How does an AI agent work?

Most agents follow the same loop, even when the details differ. First, the agent takes in the goal and the context: what you asked for, what it already knows, and what tools are available to it. Then it plans, breaking the goal into a sequence of steps and deciding which step comes first. Then it acts, calling a tool — a web search, an app, a piece of code, an online service — and watching what comes back. Then it checks the result against the goal: did that step work, or does the plan need to change? It repeats that loop until the goal is reached, it hits a limit you set, or it decides it needs to ask you something.

More advanced agents can coordinate specialised sub-agents, handing parts of a job to smaller agents that each handle one piece. Apple’s listing describes exactly this: its agents “plan and carry out multi-step tasks across apps, browsers, command lines, code, and APIs, and they can coordinate specialized sub-agents along the way.” That is also how agents can do work that would take a person a long time: one sub-agent can research, another can write, and a third can check the result, all inside the same job.

How is an agent different from a chatbot or an assistant?

The line is not always sharp, because companies use the word “agent” loosely, and many products labelled as agents are really chatbots with a few extra features. IBM’s own researchers have warned about this: what is commonly sold as an agent is often a language model with basic planning and tool-calling added on, rather than a truly autonomous system. Three questions help tell them apart. First, does it plan, or does it only respond? A chatbot answers each message as it comes; an agent works out a sequence of steps before it starts. Second, does it act, or does it only write? An agent can change things — send a message, change a file, make a booking — while a chatbot only produces text. Third, does it keep going without you? An agent should be able to pursue a goal across several steps with limited supervision, and only come back to you for decisions it cannot safely make alone.

Voice assistants such as Siri sit somewhere in between. They can carry out single commands — set a timer, send a text — but they do not plan multi-step work or choose their own tools. That is why Apple’s reported project is significant: it would move the company from an assistant that obeys commands to an agent that takes on jobs.

What is Apple actually building?

According to the reports, Apple is hiring a senior engineering manager to lead a team building “general-purpose computer-use agents”, working across the Siri, user-interface frameworks and Apple Foundation Models teams. The job listing, titled Senior Engineering Manager, Agentic AI – Proactive, was posted on 17 November 2026 and has since been taken down; Apple’s careers site now says the role cannot be found. Apple has not confirmed the project and did not respond to requests for comment from TechCrunch.

Bloomberg, which first spotted the listing, reports that Apple has eyed introducing the service as soon as the end of next year, but that the work remains in the early stages. No product name, feature list or supported hardware has been announced. The listing itself stresses safety, saying that agents able to run code and act on a user’s behalf need “strong permission models, sandboxing, and transparency.” The push comes under Apple’s new chief executive, John Ternus, who took over on 1 September 2026 and has reportedly been pressing software chief Craig Federighi to move faster on artificial intelligence.

Why is Apple also making it harder for other agents?

At the same time as building its own agent, Apple is tightening the rules for everyone else’s. In October, Apple warned developers that Full Disk Access, a permission originally meant for backup apps, was being used in ways that could expose browsing history, mail and messages without users fully understanding what they had agreed to. On 16 November, the macOS 27.3 beta arrived with the promised changes: the old Device Control and Data Access setting splits into two settings, users can limit an app to chosen folders or apps, require it to ask every time, and see an on-screen indicator when one app is steering another. Developers were told that anyone handing over this power must understand that “it puts all of the data on their Mac at risk.”

Reports say that turning on ChatGPT’s ability to control a Mac in the macOS 27.3 beta now takes 14 distinct steps, up from four. The contrast is the point: Apple wants its own agent to be able to use a computer, but it wants third-party agents to get that power only after the user has made a deliberate, informed choice. Whether that is a safety measure, a competitive advantage, or both, is already a matter of debate among developers.

What can AI agents already do today?

Agents are already working in several corners of the software world. OpenAI has launched ChatGPT for accounting teams, putting agents to work on bookkeeping tasks, and its Dots product gives its assistant its own space to carry out multi-step jobs. Meta has its Muse assistant, and Google’s agent tools are being built into its developer products. In workplaces, agents are used to triage support tickets, draft reports, check code, summarise meetings and chase up routine admin. In software development, agents can write and debug code across a whole project, run tests, and suggest fixes. None of this is science fiction; it is the market Apple is now trying to enter, late, with the advantage of owning the operating system and the hardware the agent would run on.

That lateness matters. Apple was slow to ship its first wave of AI features, and its Siri revamp has taken longer than the company planned. An agent that can genuinely use a Mac would be a bigger step than a smarter voice assistant, because it would let Apple offer the kind of hands-off help that rivals are already selling. The question is whether Apple can build one that is both capable and trustworthy, and whether it can do it before users get used to giving that job to someone else’s agent.

What are the risks?

An agent that can act on your behalf can also act wrongly on your behalf. It can misunderstand a goal, pick the wrong tool, send the wrong message, delete the wrong file, or spend money you did not mean to spend. Reporting this month has already raised the question of who carries the loss when an agent given access to money makes a costly mistake. The more tools an agent can reach, the bigger the consequences of a mistake — and the more attractive it is to attackers who want to hijack it. Researchers have shown that agents can be tricked by instructions hidden in web pages, documents or messages, and then made to act against their owner’s interests.

That is why the safety language in Apple’s listing matters. Strong permission models mean the agent should only be able to do what you have allowed it to do, and no more. Sandboxing means it should work inside limits that stop it reaching parts of your computer it does not need. Transparency means you should be able to see what it did, and why. Until those are proven in practice, the sensible approach for any user is to give an agent the narrowest access that lets it do the job, to check its work before it does anything irreversible, and to keep backups of anything important.

What happens next?

Watch three things. First, macOS 27.3, which will show how Apple balances agent power against user control when it ships. Second, whether Apple talks about its own agent publicly, and when; a launch aimed at the end of next year would put it in front of users in late 2027 if the reports are right. Third, whether the industry settles on rules for what agents may do without asking — because the agent that can use your computer is only useful if you can trust it with your computer. For now, the safest summary is the one in the job listing itself: imagine an AI that can use a computer the way a capable person does. Apple is trying to build it. It has not built it yet.