AI agent

Most agents run a loop: the model receives the goal, the available tools and the history so far; it chooses an action, such as calling an API; the system executes it and returns the result; and the model decides the next step. The ReAct pattern (Yao et al., 2022), which interleaves reasoning and actions, is an influential early description of this loop. Agents may also keep memory between sessions and hand tasks to other agents.

Why it matters: agents can complete whole workflows, such as researching leads, triaging support tickets or fixing a bug, rather than just drafting text. That also raises the stakes: an agent with access to email, payments or production systems can cause real damage if it misunderstands a task or is manipulated by malicious content. Evaluate agent products on permissions, approval steps, logging and error handling, not only on demos.

Example: a sales agent receives a list of companies, searches the web for each one, fills in fields in the CRM and drafts an introduction email for a person to approve.

Limits: agents still make mistakes that compound across steps, can loop or stall, and their cost is harder to predict because each task may take many model calls.