Tech Briefing. Your daily dose of tech news — signal, not noise.
Learning Center
One technology, explained in plain English
DTBRIEF

AI Agents, Explained: What They Are, How They Work, and Why Siri Wants to Be One

A chatbot answers you. An agent does the job. Here is the difference, the loop that makes it work, and why every big tech company is racing to build one — including the part where they still fail.

A small robot assistant thinking at a desk, choosing between tools — a browser, code, a calendar, a database and a search — while it works toward a goal
An agent is given a goal and picks its own tools to reach it. The model decides the next step — that is the whole difference.
Goal, not a prompt
An agent is given a goal and decides its own steps; a chatbot answers one message and stops
The loop
Perceive, plan, act, observe, repeat — until the job is done or it asks for help
26.4%
Share of real coding tasks the best agent fully finished in a new real-world test — the gap between demo and done is still wide

1 What it is

An AI agent is software that takes a goal and works toward it on its own. You tell it what you want done; it figures out the steps, uses tools to take them, checks the results, and keeps going until the job is finished — or until it needs your help.

The clearest definition comes from Anthropic, the company behind Claude, in its engineering guide “Building effective agents”: agents are “systems where LLMs dynamically direct their own processes and tool usage.” In plain English: the model decides what to do next. That is the whole line. If a human wrote the steps in advance, it is not an agent — it is a workflow. Anthropic calls both of them “agentic systems,” but the difference is who is in charge of the next step: your code, or the model.

Think of it this way. A chatbot is a conversation. A workflow is a recipe. An agent is a cook you tell what dinner should be — it chooses the steps, tastes as it goes, and adjusts.

2 How it works: the loop

Every agent, whatever it is built with, runs the same loop:

1. Perceive — take in the goal and the current state of things. 2. Plan — decide what to do next. 3. Act — use a tool: search the web, run code, read a file, call an app. 4. Observe — look at what the tool returned, and update the plan. Then repeat, until the goal is met or the agent hits a limit and asks a human.

Researchers call this the ReAct pattern — reason and act, in turns — from a 2022 paper by Yao and colleagues. OpenAI researcher Lilian Weng summed the ingredients as: Agent = LLM + memory + planning + tool use. The language model is the brain; the tools are the hands; the memory is what lets it keep context across steps; the loop is what turns one reply into a finished job.

A concrete example: ask an agent to fix a bug in your checkout page. It reads the code, runs the tests to see the failure, edits the code, runs the tests again, sees a new error, adjusts, and repeats — then tells you what it changed. A chatbot would have described the fix in words and stopped.

3 Agent vs chatbot vs workflow

A chatbot maps one message to one reply. It does not hold a goal across steps, it does not call tools to change anything in the world, and it forgets the task when the reply ends.

A workflow sits between them: it uses tools and can run many steps, but the path is written in advance by a developer. It is predictable and cheap — and it breaks the moment the task goes off-script.

An agent is the third thing: it holds the goal, picks its own tools and order, and adapts when the results surprise it. That freedom is the point — and the cost. Agents are slower and more expensive than a single model call, and every step is a chance for a small error to compound into a wrong turn. Anthropic’s own advice is to start with the simplest thing that works and add agency only when the flexibility is worth it.

ChatbotOne message → one reply. No tools, no memory of the task — it stops when the reply ends.
AI agentOne goal → many steps. Picks its own tools, checks the results, keeps going until the job is done.
A chatbot answers. An agent acts — and the difference is who decides the next step.

4 Why everyone’s building them now

Three stories in today’s briefing are the same story. Apple is racing to turn Siri into a real agent: fine-grained control of its own apps in iOS 27.3, expected in December, then third-party actions like booking a restaurant or ordering food in the spring — a Siri that tracks goals and keeps going without being told every step. Anthropic’s new Claude Fable 5.2 is built for agent work, and the company says a team of about 1,000 Fable 5.2 agents formalized Fermat’s Last Theorem in about 30 hours. And coding agents are the first big proving ground: they are sold as autonomous workers, not assistants.

A phone assistant rising from a smartphone and reaching out to apps — photos, calendar, restaurant, delivery and music — to carry out tasks on its own
An agent does not just answer — it reaches into your apps and does the task. That is what Apple is racing to give Siri.

The bet is simple. A chatbot can answer questions; an agent can do jobs. The company that ships the agent people trust with real tasks owns the next interface — the same way the search box and the app store did before it.

5 The honest part

Here is the catch, from today’s briefing too. A new real-world coding test, Co-Design Bench, built by Alibaba and university researchers from real user change requests, found the leading model fully completed only 26.4 percent of selected tasks — even though comparable models score above 99 percent on some older benchmarks. On average the top model met about 70 percent of the requirements; it just rarely delivered a complete, working result.

That gap is the whole story of agents in 2026: the demo is dazzling, the finished job is hard. Partial success is not a finished task, and the benchmark is new — independent replication is the next test of how durable the numbers are. But the direction is set: agents are the product every lab and platform is building toward, and the race now is not who can demo one, but who can ship one that finishes the job.

Key terms

  • Agent Software that pursues a goal by deciding its own steps and using tools, in a loop.
  • Workflow A fixed path of steps written in advance; the model may reason inside a step, but the path never changes.
  • Tool Anything the model can use to act: web search, code execution, a file system, an app’s actions.
  • ReAct loop Reason, act, observe, repeat — the pattern underneath almost every agent.
  • Context Everything the model can see at one moment: the goal, the conversation, tool results, memory.
From today’s briefing

This lesson comes from today’s Daily Tech Briefing (Episode 29, November 10, 2026): Apple’s Siri agent push, Anthropic’s Fable 5.2, and the Co-Design Bench reality check. Read the full briefing: Anthropic’s Fable 5.2 and Mythos 5.2 — Daily Tech Briefing.

Sources

Anthropic, “Building effective agents” (December 19, 2024) — the definition of agents vs workflows; Yao et al., “ReAct: Synergizing Reasoning and Acting in Language Models” (2022); Reuters on Anthropic’s Fable 5.2 launch; The Verge on Apple’s Siri agent plans; The Decoder and Precise Impact on the Co-Design Bench results.