From chat to work
A chatbot answers a question. An agent gets a task, for example “find the bug in this program and fix it”, and plans the steps itself: it reads the files, runs the tests, edits the code and checks the result. Along the way it uses tools such as a terminal, a browser or company applications.
Model Context Protocol
For an agent to use different services (calendar, email, database), it needs a common interface. The open MCP protocol, supported by most large companies, has become that standard.
Risks
- Injected instructions (prompt injection): a web page or email may contain hidden text that tries to persuade the agent to do something harmful.
- Irreversible actions: an agent should not delete files, send money or send emails without confirmation.
- Good agents therefore work with limited permissions and ask a human before important steps.