Anthropic's definition anchors AI agent explainer

The piece is built around a single organizing idea: an AI agent is not a new kind of model but an LLM placed inside a loop. The LLM reads the goal and current state, picks an action, a tool executes it, and the result goes back to the LLM for the next decision. The explainer argues that Anthropic, OpenAI's Agents SDK and the Wang et al. survey paper all describe the same four parts under different names, which is why it settles on LLM, instructions, tools and memory as the common vocabulary.

The stages of that loop — receiving the goal, decomposing it into tasks, selecting and running tools, checking results, and stopping when the completion condition is met — are illustrated with a competitive pricing comparison that ends in a Slack post. The explainer pairs each stage with a documented mechanism, such as OpenAI's guidance that an agent stops when it returns a final answer without calling a tool, or when it hits a configured max_turns limit.

The second half turns to the limits that follow from the architecture itself: errors compounding across loops, a bounded context window, wider tool permissions raising the cost of a mistake, and time and cost growing with each pass. The proposed countermeasures — guardrails, least-privilege tool permissions, human approval before irreversible actions, and stop conditions with execution logs — are presented as design choices a buyer can check for in a vendor's offering. The explainer closes by suggesting readers test one internal task against four conditions for delegability, and compare vendors on which LLM, which tools and how much permission an agent is granted.