AI agents, in plain words
AI agents, explained. Then build your own.
An AI agent works toward a goal for you: it plans the steps, uses your apps and checks its work, while you approve what matters.
- No code
- ChatGPT, Claude or Copilot
- You approve what matters
01What it is
What is an AI agent?
The short answer
An AI agent is AI that works toward a goal for you. You give it a job; it plans the steps, uses tools such as your apps, files or the web, checks what comes back, and stops when the job is done, or hands it back to you when it is stuck. You decide what it may do without asking.
“Agents are systems that independently accomplish tasks on your behalf.”
“Agents … are systems where LLMs [large language models] dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.”
Anthropic Building effective agents
“AI agents are software systems that use AI to pursue goals and complete tasks on behalf of users.”
Google Cloud What are AI agents?
Chatbot, assistant, automation or agent?
The four get mixed up all the time. The clearest difference is who decides the next step.
| Kind | What it does | Who decides the next step |
|---|---|---|
| Chatbot | What it doesAnswers your message, then waits for the next one. | Who decides the next stepPreset rules, or your next message. |
| Assistant | What it doesHelps with your task inside the conversation while you steer. | Who decides the next stepYou, one request at a time. |
| Automation | What it doesRuns the same fixed steps every time, in the order someone wrote them. | Who decides the next stepWhoever wrote the steps. |
| Agent | What it doesWorks toward a goal: picks its next step and tool, checks the result, and stops or asks when it is done or stuck. | Who decides the next stepThe AI, within the limits you set. |
Gartner's rule of thumb: use agents when decisions are needed, automation for routine workflows and assistants for simple retrieval. Gartner, June 2025
02How it works
How an agent works
Anthropic puts it plainly: agents “are typically just LLMs using tools based on environmental feedback in a loop.” Here is that loop in everyday words. Anthropic
Goal. You give it one job and say what done looks like.
Plan. It breaks the job into steps.
Act. It uses a tool: it searches, reads a file or drafts in an app.
Check. It looks at what came back and picks the next step.
Done or hand back. It stops when the goal is met or it reaches a limit, and hands back to you when it is stuck.
Plan, act, check: the loop runs until the job is done, a limit is reached or it hands the job back to you.
Before an action that is sensitive or hard to undo, such as sending, paying or deleting, a well-set agent stops and asks you first.
The seven parts of an agent
The path's first course names these seven parts, finds them in real agents from OpenAI, Anthropic and Microsoft, and has you label your own task with them.
Model
The AI that reasons and decides. OpenAI's guide calls it “the LLM powering the agent's reasoning and decision-making” (an LLM is a large language model). In the path, it is whichever model your plan gives you.
SourceOpenAI
Instructions
The job and what done looks like, then the steps, the rules, the tone and the format. OpenAI and Microsoft both advise putting rules here, not in the files you upload.
Knowledge
The reference it reads: your policies, templates and past examples. OpenAI calls it reference material, not rules or behavior.
SourceOpenAI
Tools
What lets it act: web search, files, connected apps such as your calendar or email. The model asks for a tool; the app does the work.
SourceAnthropic
Memory
What carries over between steps or chats. More isn't always better: Anthropic notes that recall gets worse as the context grows.
SourceAnthropic
Trigger
What starts a run: you, a schedule, or an event such as a new email. ChatGPT, Claude and Copilot all offer schedules, on some plans; until you set one, you start every run yourself.
Guardrails
Limits and approvals: what it must never do, and where it must stop and ask you first, before anything sensitive, irreversible or high-stakes.
SourceOpenAI
03Good for
What agents are good for at work
The jobs OpenAI, Anthropic and Microsoft suggest are everyday ones: work you repeat, from information you already have, with a result you can check.
Recurring briefs and reports
A Monday summary of last week's project notes, a daily briefing, a weekly status update.
Anthropic lists daily briefings and weekly reports among common scheduled tasks, and Microsoft offers a Status Update Agent template.
Answers from your own documents
An agent that answers policy questions from your company's documents.
Microsoft's My Company Policy agent template “answers HR policy, holiday, and PTO questions” from company documents.
SourceMicrosoft
Sorting and draft replies
When a message that matches your rule arrives, it sorts it and drafts a reply for you to send.
ChatGPT Work can respond when a new Gmail or Slack message arrives, and Copilot Cowork can start a task when a matching email arrives.
Research you repeat
Tracking a topic, a market or your suppliers, and writing up what changed.
Anthropic lists recurring research among common scheduled tasks; OpenAI says ChatGPT Work can research a topic and analyze information.
Documents in your house style
Emails, briefs and slides that follow your templates and brand rules.
Anthropic's examples of custom skills include applying brand style guidelines and following company email templates.
SourceAnthropic
Paperwork from a folder
Turning a folder of receipts into a formatted expense report.
One of Anthropic's own examples for Claude Cowork.
SourceAnthropic
When not to use an agent
OpenAI's guide names three signs that an agent beats a fixed set of rules: complex decisions, rules that are hard to maintain, and a heavy reliance on unstructured information. Otherwise, in OpenAI's words, “a deterministic solution may suffice”: fixed rules that give the same result every time. OpenAI
- A plain prompt
One good request does the job. Anthropic's advice is to find “the simplest solution possible” first.
- An automation
The steps never change. A fixed workflow skips the extra cost and waiting that an agent's reasoning brings.
- A person
A mistake would be costly and you can't check or undo it. OpenAI says actions like these should trigger human oversight.
- An agent
The job needs judgement calls, the inputs are messy, and you can check the result.
Agents at work, in three numbers
40%
of respondents from large organizations (annual revenue over $1 billion) report scaling AI agents, that is, rolling them out across the organization.
39%
of respondents expect AI-related declines in their organization's total employment in the coming year.
50%
of AI users named quality control of AI output as a human skill that matters more as AI takes on more work: the top answer.
Each figure answers its own survey's question: read them as context, not as one trend.
04Build one
Build one in ChatGPT, Claude or Copilot
All three offer the same building blocks: instructions, reference files, skills, connected apps, schedules and approvals. The names and the plans differ. Here is what each tool calls them, at the level of free plan, paid plan or work license; the courses give the exact steps.
ChatGPT OpenAI
Free plan to start- Instructions and files
- A project keeps your instructions, files and chats together, on every plan, the free one included.
- Skills
- A skill is a reusable workflow ChatGPT uses when it fits. Bundle skills and apps into a plugin by describing it in a chat; OpenAI's pricing page lists creating plugins on its paid personal plans, not the free one.
- Your apps
- Apps connect ChatGPT to services such as Google Drive or Slack, and you choose how much each may do without asking.
- On a schedule
- Scheduled tasks: ask in plain words and ChatGPT runs the task again later. How often depends on your plan.
- When something happens
- ChatGPT Work, OpenAI's agent for longer jobs, can respond to new Gmail or Slack messages and GitHub changes, on most paid plans.
- Approvals
- OpenAI says letting an app act without asking “carries elevated risk”. A scheduled task pauses when an action needs your approval.
What about custom GPTs? See the answer below
SourcesProjectsSkillsPluginsPricingAppsApp permissionsScheduled tasksChatGPT Work
Claude Anthropic
Free plan to start- Instructions and files
- A project has its own chats and knowledge base. Anthropic makes projects available on every plan, the free one included.
- Skills
- A skill is a folder with a SKILL.md file that teaches Claude one procedure. Skills need code execution switched on.
- Your apps
- Connectors link Claude to apps such as your calendar and documents. For each tool you choose whether it runs, asks first or is blocked.
- On a schedule
- Scheduled tasks, on paid plans. They run in the cloud, even when your computer is asleep, unless a task needs files or apps on your computer.
- When something happens
- Anthropic's help page lists time-based schedules only, so the path's Claude track builds a scheduled check, on a paid plan: each run reads a window that matches its cadence, such as the past 24 hours, and checks the condition first. Claude Code, a developer tool the path doesn't teach, has event triggers of its own.
- Approvals
- Cowork, Claude's agent for multi-step work on paid plans, has approval modes; with approvals skipped, Anthropic warns, “nothing checks its actions.” It asks before permanently deleting files.
SourcesProjectsSkillsConnectorsScheduled tasksClaude CodeCoworkCowork safety
Copilot Microsoft
Work or school account- Instructions and files
- Agent Builder in Copilot Chat: describe the agent in plain words or write its instructions. Without a Copilot license, Microsoft says agents can be grounded on instructions or the web; files you upload and SharePoint need a Copilot license or pay-as-you-go billing, and email and Teams need a license.
- Skills
- A skill is a folder with a SKILL.md file. For skills in an Agent Builder agent, Microsoft asks for a Microsoft 365 subscription with access to Copilot Chat, and says a Copilot add-on license isn't required.
- Your apps
- Agent Builder agents don't take actions in other services. For actions, Microsoft points you to Copilot Studio.
- On a schedule
- Copilot Cowork runs scheduled tasks. It needs a Copilot license and usage-based billing, and your admin turns it on.
- When something happens
- Cowork can start a task when a matching email arrives or a Teams message is posted.
- Approvals
- Microsoft says Cowork “doesn't take actions without your approval”, though you can choose to skip the prompt for similar actions.
Microsoft Copilot was formerly named Microsoft 365 Copilot. Microsoft documents agent building for work and school accounts, and your organization's admin can switch Agent Builder off. With a personal Microsoft account, take the ChatGPT or Claude track.
SourcesAgent BuilderLicensesSkillsCowork setupCowork FAQCopilot's name
Some of what you build can travel
Two open standards sit under these tools, so parts of what you build, such as a skill folder or a connection to an app, can work in more than one of them.
MCP Model Context Protocol
“An open-source standard for connecting AI applications to external systems.” Its own docs call it “a USB-C port for AI applications.” Anthropic launched it in November 2024 and donated it in December 2025 to the Agentic AI Foundation, under the Linux Foundation. You meet it as a connector or an app; you don't have to build one.
SourceAgent Skills The SKILL.md folder
“A skill is a folder containing a SKILL.md file that teaches an agent a specific task or domain.” Anthropic published the standard in December 2025 and gave it to the same foundation in October 2026; it says more than 70 agent platforms, ChatGPT among them, have integrated it. Microsoft's documentation describes Copilot skills the same way.
Source
05The path
The path: from first idea to a working agent
5 steps, one agent. In the first course you pick one real task from your work; every course after it takes that same agent, and its Agent Card, a step further. The lessons teach each building block once, then show it in ChatGPT, Claude and Copilot.
Step 1 Agent Card · v0
Understand agents
What agents are and aren't, how the loop works, what can go wrong, and whether your task needs an agent at all.
You leave withAgent Card v0 for one real task of yours: a fit verdict, its risks and approvals, and the tool and plan you will build it in.
Step 2 Agent Card · v1
Build version 1
Write instructions that steer, keep rules apart from reference files, add examples and edge cases, and test before you trust.
You leave withAgent v1 in your tool, with structured instructions, reference files and a first test sheet. By Microsoft's definition it is an agent; by OpenAI's stricter line, not yet: it answers when you start it, and doesn't act in your apps until step 3 connects one.
Step 3 Agent Card · v2
Give it skills and tools
Teach it one of your procedures as a skill, connect one app with reading allowed and writing set to ask, and rate every tool's risk.
You leave withAgent v2: one skill, one connected app and a tool risk table.
Step 4 Agent Card · v3
Test and safeguard it
Build a test set from real work, run it more than once, decide what the agent may do alone, and test it against hidden instructions.
You leave withTest results, an approval map, a prompt-injection test with its fix, and a one-page safety case.
Step 5 Agent Card · v4
Put it to work
Let it run on a schedule or when something happens, write a stop plan and test the off switch, review its runs and costs, and hand it to your team.
You leave withA shipped agent that does one recurring job, checked against twelve acceptance criteria, with its stop plan, run log and hand-off note.
5 courses 40 lessons 11 hr in all
Lessons 1 and 2 of each course are free with a free account. The rest of each course and its certificate come with Pro. Pro from $4.92/mo, billed yearly. See what Pro includes
One Agent Card, start to finish
Everything you decide about your agent goes on one page of plain text: its job, when it runs, which tools it may use and at what risk, when it must ask you, and what it never does. You start it in the first course and update it in every course after. Because it is plain text, it moves with you if you change tools.
- v0Understand
- v1Build
- v2Add tools
- v3Test
- v4Ship
- Name
- Monday Brief
- Job
- Turn last week's project notes into a one-page brief for the team.
- When it runs
- Mondays at 8:00
- Tools
- Project folder: READ, low risk, your account · Email to the team: WRITE, high risk, your account
- Knowledge
- The brief template and the last four briefs
- Done when
- Every item in the brief points to a note
- Hand back when
- A note is missing, or two notes disagree
- Needs approval
- Anything sent to people
- Never does
- Delete or move files
- Owner
- You
One Monday run
- Read last week's six project notes.
- Drafted the one-page brief from the template.
- Stopped and asked before emailing it to the team.
06Who it's for
Who this path is for
It's for you if
- You already use ChatGPT, Claude or Copilot at work and keep hearing about agents.
- You lead a team and need to judge whether an agent is worth trying.
- You have one repeatable task you would like to hand off, and you want to stay in charge of it.
It is not a coding course. If you want to build agents with APIs and code, the ideas carry over; the code is not here.
What you need
- A ChatGPT or Claude account: a free plan is enough to start. Letting an agent run on a schedule may need a paid plan, and the lessons show a fallback when yours doesn't have it.
- For the Copilot track, a work or school Microsoft account.
- One real task from your work. No code and no developer tools.
- If you will connect work accounts, check your company's AI policy first.
How long it takes
The 5 courses come to 40 lessons, about 11 hr in all. Each course ends with a capstone built on your own task.
07Risks
Limits and risks, plainly
Agents are useful because they act. That is also the risk. Here is what goes wrong, with real cases, and the habits that keep you in charge.
Prompt injection
Hidden instructions
An agent reads web pages, emails and files, and any of them can carry instructions aimed at it. OpenAI's own example: a malicious comment the agent comes across that directs it to retrieve a password reset code from Gmail and send it to a malicious website. OpenAI says prompt injection is “unlikely to ever be fully ‘solved’”, and in NIST's January 2025 tests of an agent built on a 2024 model, the strongest new attack succeeded 81% of the time, against 11% for the strongest baseline attack.
SourcesOpenAI Help CenterOpenAINIST
Wrong actions
Too much access, too little checking
OWASP names three root causes of what it calls excessive agency: excessive functionality, excessive permissions and excessive autonomy. Anthropic also warns that an agent's autonomy brings “the potential for compounding errors.”
Cost
More steps, more usage
Anthropic measured agents using about four times the tokens of a chat, and multi-agent systems about fifteen times; tokens are the small units of text, such as words or parts of words, that a model reads and writes. Multi-step work uses more of your plan than a quick question, and on business plans agent runs can draw on credits.
SourcesAnthropicAnthropic glossaryClaude Help CenterOpenAIMicrosoft
It has happened
June 2025
EchoLeak
A zero-click prompt injection in Microsoft 365 Copilot: a single crafted email could get data out of it, without anyone clicking anything. It is recorded as security flaw CVE-2025-32711.
SourcearXiv
June 2025
Project Vend
Anthropic reported that in its own experiment, an agent running a small office shop told customers, for a time, to pay into an account it had made up.
SourceAnthropic
May 2025
GitHub MCP
Researchers showed that a malicious issue, a public post on a GitHub project, could hijack a user's agent and make it leak data from the user's private code repositories.
SourceInvariant Labs
July 2025
Replit
Tom's Hardware reported that Replit's coding agent wiped a company's live production database during a code freeze, when no changes were allowed, despite instructions saying “NO MORE CHANGES without explicit permission.”
SourceTom's Hardware
How to stay in charge
- Give it one narrow job: OpenAI says it's “safer to ask your agent to do specific things.”
- Give it only the access the job needs, as OpenAI advises: “only the sensitive data or credentials it needs to complete the task.”
- Keep approval on for anything it can't undo: sending, paying, deleting.
- Read each approval before you click it: check that the action looks right and that what it shares is fine to share.
- Test it more than once. Anthropic's arithmetic: an agent that succeeds 75% of the time passes three runs in a row only about 42% of the time.
- Know how to pause it and switch it off, and set limits on retries and actions.
SourcesOpenAI, prompt injectionsAnthropic, agent evalsOpenAI, building agents
08FAQ
Questions people ask about AI agents
What is an AI agent?
AI that works toward a goal for you. It plans the steps, uses tools such as your apps or the web, checks the results, and stops or asks you when it is done or stuck. OpenAI's definition is short: “Agents are systems that independently accomplish tasks on your behalf.”
What's the difference between a chatbot, an AI assistant and an AI agent?
A chatbot answers; an assistant helps while you steer; an agent chooses its next step and its tool until the goal is met. Google Cloud ranks them by autonomy, with bots the least autonomous and assistants needing your input and direction. OpenAI says simple chatbots are not agents.
Is an AI agent the same as an automation like a Zapier workflow?
No. A workflow follows steps someone wrote in advance; an agent decides its steps as it goes. Gartner's advice is to use agents when decisions are needed, automation for routine workflows and assistants for simple retrieval.
Do I need to know how to code to build one?
No. ChatGPT projects and plugins, Claude projects, skills and connectors, and Microsoft's Agent Builder are all set up by writing instructions and describing what you want in plain words.
Can I build an AI agent for free?
A simple one, yes. On the free plans of ChatGPT and Claude you can set up a project with your files, give it your instructions and connect apps, which is where the path starts. Microsoft says Copilot Chat users have access to agents grounded on instructions or the web, but its own documentation also says users without a Copilot license don't have agent authoring, so ask your admin before you count on it. Letting an agent run on its own, on a schedule or when something happens, mostly needs a paid plan or a work license.
Which is best for agents: ChatGPT, Claude or Copilot?
The one where your work already lives. Building your own agent in Copilot needs a work or school account, and Microsoft says agents grounded in your organization's data need a Copilot license or pay-as-you-go billing. ChatGPT and Claude work with personal plans and connect to apps such as Google Drive, Slack or your calendar.
What happened to custom GPTs?
OpenAI stopped new GPT creation on personal plans and set December 11, 2026 as the standard retirement date for custom GPTs (February 11, 2027 for some Enterprise and FedRAMP workspaces). A GPT can migrate to a plugin, where its instructions become a skill, and OpenAI suggests a project as the simple replacement for a GPT that kept instructions and files together.
What is MCP, and do I need to understand it?
The Model Context Protocol is an open-source standard for connecting AI applications to other systems; its own docs compare it to a USB-C port for AI applications. Anthropic launched it in 2024 and donated it to the Agentic AI Foundation, under the Linux Foundation, in December 2025. You meet it as a connector or an app in your AI tool, so you don't need to build one.
Are AI agents safe? What is prompt injection?
Prompt injection is when instructions hidden in something the agent reads, such as a web page, an email or a file, steer what it does. OpenAI says it is “unlikely to ever be fully ‘solved’”, so keep each job narrow, give the agent only the access it needs, and read every approval before you click it.
Will an agent act without asking me?
Only as far as you let it. ChatGPT and Claude let you choose, app by app or tool by tool, what may run without asking; ChatGPT's workspace agents ask before write actions by default, and Microsoft says Copilot Cowork doesn't take actions without your approval, though it lets you skip the prompt for similar actions. Be careful with settings that skip approvals: Anthropic warns that with them, nothing checks the agent's actions.
How much does it cost to run an agent?
More than chatting. Anthropic measured agents using about four times the tokens of a chat (tokens are the small units of text a model reads and writes). Multi-step work uses more of your plan's allowance than a quick question, and on business plans agent runs can draw on credits.
Will AI agents replace my job?
Nobody can say for certain. In McKinsey's State of AI 2026 survey, 39% of respondents expected AI-related declines in their organization's total employment in the coming year. In Microsoft's 2026 Work Trend Index, quality control of AI output was the top answer (50%) when AI users were asked which human skills matter more as AI takes on more work, and that is the skill this path practices: directing an agent and checking its work.
Keep learning
Go deeper
What changed in AI robots, checked daily, with a source behind every entry: AI Robots Tracker
Start with one real task.
Bring one job you would like off your plate. The first step turns it into a plan you can build.
Start lesson 1 freeProduct facts on this page were checked against OpenAI, Anthropic and Microsoft documentation on November 13, 2026.
