The Agent as a System

When most people hear the phrase 'AI agent,' they picture a very smart chatbot — a language model that answers questions. That picture is incomplete. A modern AI agent is a system: a collection of distinct components that work together to perceive a goal, plan a course of action, invoke external capabilities, remember what it has done, and iterate until the goal is met. The language model sits at the center of that system, but the system is far more than the model alone.

Consider a concrete example. You ask an AI research agent: 'Find the three most-cited papers on transformer attention published in the last two years, summarize each one, and save the summaries to a file.' To complete that task the system must search a database, read full-text documents, synthesize information across multiple sources, write formatted summaries, and call a file-write API. No single forward pass through a language model can do all of that. Instead, a loop of reasoning, tool calls, and memory operations runs repeatedly until every sub-task is complete.

Five Components of the Modern Agent Stack

Researchers and engineers have converged on a rough consensus about what a general-purpose agent contains. The five canonical components are: (1) the LLM reasoning core, which does the language understanding and generation; (2) the planner, which decomposes a high-level goal into an ordered sequence of concrete actions; (3) the tool layer, which connects the agent to external capabilities like search, code execution, databases, and APIs; (4) the memory system, which stores context, past actions, retrieved facts, and long-term knowledge; and (5) the orchestration loop, which drives the entire cycle — prompting the model, executing the action it chooses, feeding results back in, and repeating until done. Each of these five components has its own architecture, failure modes, and design tradeoffs. Modules H1-02 through H1-06 examine each in turn.

Why the Systems View Matters

Understanding agents as systems rather than as clever prompts is practically important because it changes where you look when something goes wrong and what you change when you need improvement. If a research agent produces outdated summaries, the fault might be in the retrieval tool (wrong database, stale index), the memory system (results from a previous task bleeding in), the planner (wrong decomposition of the search step), or the model itself (poor synthesis). Treating the whole thing as a black-box model makes diagnosis impossible. Treating it as a system with distinct components makes every failure traceable.

This systems perspective also clarifies capability boundaries. An LLM on its own has no ability to browse the web, execute code, or write a file — it can only generate text. An agent with a web-search tool and a code-execution sandbox can autonomously research and verify. The same model, radically different capability, achieved entirely through architecture rather than through changes to the model weights. This is why agent engineering has become a distinct and valuable discipline alongside model training.