Agentic AI

Anthropic says the word agent is used in several ways. Some people mean a system that works on its own for a long time. Others mean a more scripted setup. Anthropic calls the whole family agentic systems, and then splits it¹. A workflow orchestrates a model and tools through code paths that are fixed in advance. An agent is a system whose model directs its own process and tool use¹.

In the agent pattern, the model plans, calls tools, reads the result and can continue for many steps. It may pause for a person at a checkpoint, and the run can stop after a set number of iterations¹. The basic tools include search and calls to external services¹. Anthropic's coding example edits many files from a task description and can use test results as feedback¹.

Anthropic says these systems often trade latency and cost for better task performance. The more open-ended agents also cost more, and errors can compound across steps. The advice is to use the simplest design that works: a workflow when the task is well defined, and an agent only when the number of steps cannot be fixed in advance¹. OpenAI's guide to agents, which does not use the word agentic, says high-risk actions such as cancelling an order, a large refund or a payment should trigger human oversight². Its API can pause a tool call until a person approves a side effect such as a cancellation or an edit³.

Anthropic points to customer support and coding as uses. Support tools can pull order history or issue a refund, and coding agents can check their work against tests¹. The same post says both uses depend on human oversight¹.