Agent Core Capabilities: A Full Breakdown from Interview Question to Agent Architecture

Interviewers often ask “What are the core capabilities of an Agent?” A brief answer such as “it can call a large model” is insufficient; the response must outline the underlying architecture.

Core capabilities of an Agent

Perception – Converts raw text, image, audio, or sensor signals into structured understanding. Example: an autonomous car recognizing a red light as the semantic need to stop.

Planning – Decomposes complex intents into executable tasks and can re‑plan when obstacles appear. Example: a travel planner turning “arrange a northern‑lights trip” into visas, flights, locations, and clothing.

Memory – Short‑term memory maintains conversational coherence (e.g., remembering a statement from two minutes ago). Long‑term memory stores preferences, history, decisions, and knowledge (e.g., recalling a user’s dislike for sweet drinks and recommending an unsweetened coffee). The purpose is to retrieve the right information at the right moment.

Tool use – Extends the Agent beyond its internal knowledge by invoking external services such as search engines, calculators, databases, or APIs. Example: fetching a stock price via a financial API or drafting and sending an email.

Action – Executes the final step in the digital or physical world: sending a reply, completing a transaction, or driving a robotic arm. Action must close the loop with perception by confirming the outcome.

Reflection & learning – Analyzes execution traces, extracts improved strategies, and writes correction signals into long‑term memory so that the same mistake is not repeated.

Closed‑loop cooperation – The capabilities form a feedback‑driven cycle: perception feeds memory, planning breaks down tasks, tool use selects resources, action delivers results, and feedback returns to memory. Any broken link degrades the entire system, similar to a watch missing a gear.