AI agents
An **AI agent** (in today’s product sense) is a system that **loops**: observe → decide → act with tools → check → repeat toward a goal. A plain chatbot answers once. An agent may browse, call APIs, write files, or ask you for confirmati...
What it is
An AI agent (in today’s product sense) is a system that loops: observe → decide → act with tools → check → repeat toward a goal. A plain chatbot answers once. An agent may browse, call APIs, write files, or ask you for confirmation along the way.
Why it matters
Agents multiply usefulness and risk. They can finish multi-step work—and they can take wrong actions confidently if tools are powerful and oversight is weak (Course 29 security literacy).
How it works (plain)
Typical pieces:
- A language model as the “planner/brain”
- Tools (search, calendar, code runner, email—depends on product)
- Memory (chat history + optional long-term stores)
- Policies: when to stop, when to ask a human
Everyday example
“Check my email for invoices due this week and draft reminders” is an agent-shaped task. “What is an invoice?” is a chat question.
Try it
Write a goal in one sentence, then list the tools a safe agent would need—and which steps must stay human-approved (sending money, deleting data, public posts).
Myths
- ⚠️ Myth: Agents are autonomous people.
- ✓ Reality: They are engineered loops with models + tools + guardrails.
- ⚠️ Myth: More tools always help.
- ✓ Reality: More tools expand blast radius when instructions go wrong.
Sources
- OpenAI Academy — Agents and Workflows: https://academy.openai.com/en ↗
- ANN live: https://www.ainerdnetwork.com/learn/agents ↗
- NIST AI RMF (govern/map/measure/manage framing): https://www.nist.gov/itl/ai-risk-management-framework ↗
