The Short Version
Most AI you've used is a chat: you type, it answers. An agent is the next step — AI that can not only answer, but act. Both run on the same kind of model underneath. The difference is what the software around the model is allowed to do.
Plain Chat — the advisor on the phone
Like calling a brilliant advisor who can only talk. They answer questions, draft text, and explain ideas — but they can't touch your computer, read your files, or send anything. Every document, detail, and follow-up has to pass through you.
Agent — the co-worker at a computer
Like delegating to a co-worker with their own login. Give them a goal and they can read the files, look things up, take the actions, and come back with the finished work — checking with you at the important moments.
The Same Task, Two Ways
Take one realistic request: "Summarize the three reports my team sent this week and email the summary to everyone." Watch how differently it goes:
- You find the three reports and open each one.
- You copy the text (or upload files) into the chat, because the AI can't see your computer.
- You prompt it, read the summary, and copy the result out.
- You open your email, paste the summary in, address it, and hit send.
The AI wrote one part. You did the legwork around it.
- You give the goal once — "summarize this week's three reports and email the team."
- The agent reads the reports itself, drafts the summary, and writes the email.
- It shows you the draft for a 30-second review before anything is sent. You still approve.
- With your OK, it sends the email and reports back what it did.
You reviewed the result. The agent did the legwork.
What Makes It an Agent?
"Agent" isn't a marketing word for a smarter chatbot. It describes software with specific capabilities added around the model. The big four:
Tools — it can act, not just answer
A plain chat produces text. An agent is connected to real systems: it can search the web, read and write files, send messages, update a calendar, or query a database — through standard connections like APIs and MCP.
Multi-step autonomy — it plans and chains
You give an agent a goal, not a script. It breaks the goal into steps, does them in order, handles what comes back, and keeps going — instead of stopping after every message for you to drive.
Memory — it keeps the important parts
A plain chat forgets everything the moment the window closes. An agent keeps a short-term "whiteboard" for the current conversation and a long-term "filing cabinet" of your preferences, projects, and rules.
Verification — it shows its work
A good agent checks its own output against sources, tells you what it did, and asks before acting on anything sensitive. That approval step — plus your review — is what makes delegation safe.
This is the same chain drawn in the Glossary's "How It All Fits Together" diagram: your prompt goes in, the agent gathers context, calls the model to reason, acts through its tools, and verifies — and only then delivers the outcome.
A Spectrum, Not a Switch
This isn't "chat bad, agent good." Plain chat is still the right tool plenty of the time — the skill is knowing which one the task calls for:
Plain chat is right for…
- Quick questions and explanations
- Drafting and rewriting text you'll finish yourself
- Brainstorming and thinking out loud
- Anything where you don't want it touching your systems
Delegate to an agent when…
- The task spans files, tools, or multiple steps
- It's repetitive and you'll do it again next week
- The work happens across systems (email, calendar, docs)
- You want a checked, finished result — not just a draft
More capability, more care. An agent that can send emails and move files can also make mistakes at that scale. The rule from the Playbook applies double here: trust but verify — review what it did, keep approval gates on anything sensitive, and start with small delegations.