Short answer
AI agents are systems that use models to plan and take actions via tools—browsing, coding, updating records—toward a goal, ideally with guardrails and human oversight.
Why it matters
Agents can automate multi-step work, but unreliable reasoning plus privileged tools can cause costly mistakes without approvals and monitoring.
Diagram
Article
Ops
Track successful task completion and costly failure rates separately.
Examples
Research assistant with citation checks
An agent receives a question, searches an approved document corpus, opens top passages, drafts an answer with quotes, and stops for human approval before sending externally. Tool access is read-only; the agent cannot modify source documents.
Software triage agent
In a controlled staging environment, an agent reads failing test logs, proposes a patch, runs the test suite, and opens a draft pull request. Merge rights remain with human engineers; the agent’s write scope is limited to a feature branch.
FAQ
How is an AI agent different from a chatbot?
Chatbots primarily generate responses in conversation. Agents are built to take actions toward a goal across steps—calling tools, updating state, and continuing until completion criteria are met. Many products blend both patterns.
What is prompt injection for agents?
Prompt injection occurs when untrusted content (a webpage, PDF, or email) contains instructions that the model mistakenly treats as authoritative system direction—potentially causing unsafe tool calls. Mitigations include privilege separation, allowlists, content provenance, and requiring confirmation for sensitive actions.
Related glossary terms
- AI agent — A system that pursues goals by choosing actions over multiple steps, often using tools, memory, and feedback from an…
- Prompt — The input text or structured instructions given to a generative model to steer its output toward a desired task or…
- Reinforcement learning — Learning to choose actions in an environment to maximize cumulative reward through interaction rather than fixed labels.
Sources
-
OpenAI platform documentation (tool use / assistants concepts)
OpenAIOfficial conceptual documentation for model tool use, a core building block of LLM agents.
-
AI Risk Management Framework
NISTRisk framing for autonomous and semi-autonomous AI system deployments.
-
Google AI Blog
GoogleEducational communications on agentic and tool-using AI systems.