What is an Agent? — Anthropic's Definition

Key Takeaways

• An agent, per Anthropic, is a system where the LLM itself directs multi-step actions — choosing tools, interpreting results, and deciding next steps — rather than passively responding to a fixed prompt pipeline.
• The defining characteristic is model-directed control flow: the model reads environment state (tool outputs, API responses, context window updates) and emits the next action, closing the perception-action loop autonomously.
• In production customer-support deployments, this distinction is operationally critical — an agent can autonomously route an Pravara Bank complaint, trigger a refund API call, and update a CRM record in one loop; a chatbot cannot without hard-coded orchestration.
• Memory anchor — think of an agent as the difference between an API gateway that routes passively and a service-mesh sidecar that observes live traffic state and enforces dynamic policy: the model is the sidecar, not the gateway.