What happens between your request and the agent's result?
The agent runs a loop. It reads the goal, decides the next step, and uses a tool to take it: a web search, opening a page, filling in a form, sending an email, running a little code. It looks at what came back and decides again. It stops when it judges the goal met or when it gets stuck, and nobody approves each step unless the agent was built to ask.
Anthropic's December 2024 guide, "Building effective agents", draws a useful line. A workflow follows steps a developer laid down in advance. An agent directs its own process and picks its own tools as it goes. Plenty of products sold as agents are workflows, and for a small business that is often the safer design.
What does an agent's to-do list look like for a removalist? (example figures)
A removalist in Orange has a new 6-tonne truck and gives an agent this job: "Get three quotes for wrapping the truck in our logo, under $4,000, and accept the best one." The agent has a browser, the business email account and permission to send.
| Step | Tool | What can go wrong |
|---|---|---|
| 1. Find signwriters near Orange | Browser | It leans on the firms that paid for the top of the page, one of them three hours away |
| 2. Send the brief to six of them | Your phone number and logo now sit in six inboxes | |
| 3. Compare the three replies | It ranks $3,600 below $3,900 and misses the $900 design fee in the first quote's fine print | |
| 4. Accept the cheapest | The business is committed in writing to $4,500, over budget |
Steps 1 to 3 are chores, and an agent can save an afternoon on them. Step 4 is a decision with money attached, and it is the one to keep. The sensible version of this agent stops after step 3 and shows its working.
Which jobs should an agent never be trusted with alone?
Anything that spends money, changes a live account, publishes in your name or sends messages to customers. Those are the places where a mistake is expensive and hard to take back. OWASP, the open security project, lists "excessive agency" among the top risks for applications built on large language models: giving an AI more permissions, tools or independence than its job needs.
There is a second reason. An agent reads web pages and emails as it works, and text in them can be written to steer it, which is prompt injection. An agent holding a payment card and a browser is exposed to whatever the browser finds.
Where does CoraVision stop short of being an agent?
CoraVision is built to stop short of acting: it reads your accounts and leaves them exactly as they were, with nothing created, changed, paused or deleted, and it cannot post to your pages. Cora sees your screen during a session you start and, in pill mode, points at the next step while you take it. Every change is yours to make. The side-by-side of chatbots, copilots, agents and an AI marketing team is on chatbot, copilot, agent or team.
Questions
Can ChatGPT act as an agent?
Mostly it is a chatbot, but in July 2025 OpenAI added an agent mode that can browse websites, fill in forms and work through a task on its own, asking before some actions. The same app can be either, depending on which mode you switch on.
What does agentic AI mean?
It is the adjective, used for AI that acts with some independence. In marketing copy it stretches from a chatbot with one tool to a fully independent agent, so ask two questions: which tools can it use, and does it ask before it acts?