AI agent

An AI system that pursues a goal over many steps by planning, using tools and checking its own results.

• A language model runs in a loop: choose an action, call a tool (run code, search, edit a file), read the result, then decide again.
• Software engineering is the leading use because code gives fast, objective feedback: the tests pass or fail.
• Reliability compounds: 95% success per step gives only about 36% over 20 steps, so long tasks demand extremely reliable steps.
• Reading untrusted content creates 'prompt injection' risk: hidden text on a page can pose as instructions and hijack the agent.