What is an AI agent?

Full explainer: What are AI agents? covers this in more depth, with examples.

What it means at work

A chatbot replies to a message. An agent is given an objective and works toward it: it plans, calls tools like a spreadsheet, browser or email account, looks at the result and adjusts. Anthropic’s engineering guide defines agents as systems where LLMs dynamically direct their own processes and tool usage, in contrast to workflows with a fixed path. The useful mental model is a fast, literal-minded assistant. It can do a lot of legwork, but it still needs a clear brief, limited permissions and a human who approves what matters.

What the sources say

• Where a person fits: Anthropic says agents can pause for human feedback at checkpoints or when encountering blockers.
• Cost of autonomy: Anthropic notes that the autonomous nature of agents means higher costs, and the potential for compounding errors. A wrong early step carries into every later one.
• Testing: Anthropic recommends extensive testing in sandboxed environments, along with the appropriate guardrails.
• Trade-off: Anthropic says agentic systems often trade latency and cost for better task performance.
• Limits in practice: locally, Codex uses a sandbox plus an approval policy that controls when it must stop and ask you before acting.
• Permissions: Claude Code documents that a session’s permission mode sets which actions Claude can take without asking you first.
• Research basis: the ReAct paper explores using LLMs to generate both reasoning traces and task-specific actions in an interleaved manner.

An example

You ask an agent to collect this week’s sales exports from a folder, combine them, flag anything that changed by more than 10 percent, and draft a summary. It opens the files, does the math, writes the draft and stops for you to review before anything is sent.

A common misconception

The hype suggests agents run whole jobs unattended. In practice, the reliable pattern keeps a person in the loop at the risky steps: sending, deleting, paying or publishing. Start with read-only access and tasks where a mistake is easy to spot and undo.

Sources

• Anthropic, Building effective agents: “systems where LLMs dynamically direct their own processes and tool usage”
• Anthropic, Building effective agents: “Agents can then pause for human feedback at checkpoints or when encountering blockers”
• Anthropic, Building effective agents: “The autonomous nature of agents means higher costs, and the potential for compounding errors”
• Anthropic, Building effective agents: “We recommend extensive testing in sandboxed environments, along with the appropriate guardrails”
• Anthropic, Building effective agents: “Agentic systems often trade latency and cost for better task performance”
• OpenAI, Codex, Agent approvals and security: “Locally, Codex uses … plus an approval policy that controls when it must stop and ask you before acting”
• Anthropic, Claude Code Docs, Security: “permission mode sets which actions Claude can take without asking you first”
• arXiv, Yao et al., ReAct: Synergizing Reasoning and Acting in Language Models: “generate both reasoning traces and task-specific actions in an interleaved manner”